Changelog

2.5.2 (unreleased)

  • Enabled SSM triggers and HasSSM source property

  • Rework determining the preferred external event using a keyfunc method similar to the superevent manager. This now prioritizes, in order, by whether there exists a previous event, whether the event is likely real, whether this event has passed publishing threshold, whether the search is prioritized, whether the space-time joint FAR is better, and whether the temporal joint FAR is better.

  • Check whether MOC to assign proper filename when downloading external sky maps.

  • Launch Fermi sky map requests only after ten minutes following the FINAL notice. This more closely matches the availability of these sky maps and prevents a significant number of failed calls, both retryable and those that will fail due to no sky map being available (e.g. after a FLIGHT notice with no FINAL notice) compared to the previous method.

  • When creating combined sky maps, set any negative values in the flattened external sky map to zero. This at least prevents erroneous results when we receive a sky map with these invalid values.

2.5.1 “Cactus cat” (2024-08-20)

  • Update trials factor to account for MLy and SPIIR joining O4b.

  • Fix bug related to pesummary task failing due to recently changed naming convention of calibration filenames.

2.5.0 “Cactus cat” (2024-08-08)

  • Update ligo-rrt-chat` to v0.1.5, ligo-followup-advocate to v1.2.10, and ligo.skymap to v2.0.1.

  • Change functions calls to prepare for NumPy 2.0 release.

  • Drop support for Python 3.9 and switch from Python 3.9 to Python 3.11 for deployments for all tiers. This is to remain up to date with version support policies of Numpy, Astropy, and ligo.skymap, some of which have already dropped support for Python 3.9.

  • Add a GitLab issue template to include new external pipelines to RAVEN.

  • Switch to using igwn-alert client’s listener, since most of functionality used in our implementation is now present in the upstream client.

  • Use integer times for creating caches for omegascans, since too much precision results in problems in loading in TimeSeries.

  • Fix bug where FloatingPointErrors in omegascans would not create the fig variable, which is later referenced.

  • Switch external event listeners to GCN Classic over Kafka.

  • Move CVMFS auth token retrieval from our crontab to renew-cert.sh to fix a race condition that caused authenticated GraceDB operations to fail in production.

  • Require specific format for all GraceDB event and superevent IDs in the dashboard.

2.4.4 “Orabou” (2024-06-14)

  • Fix bug in raven.py that prevented EM_COINC being applied for targeted subthreshold GRBs.

2.4.3 “Orabou” (2024-06-05)

  • Update to ligo-followup-advocate v1.2.9.

  • Revert to custom IGWN alert that was in use until long-term stability with the new client is demonstrated.

2.4.2 “Orabou” (2024-06-04)

  • Switch to using igwn-alert client’s listener, since most of functionality used in our implementation is now present in the upstream client.

  • Bump bilby_pipe to 1.3.2 to fix calibration file lookups

  • Add search with subthreshold GRB events for CWB-BBH GW Superevents.

2.4.1 “Orabou” (2024-05-30)

  • Add GWSkyNet 2.4.1 to avoid installing tensorflow-cpu >= 2.16.1. Clairfy the version of annotated skymaps and quantities in the uploads. Add public tag to GWSkyNet’s uploads.

  • Reduce Burst trials by one removing MLy from trials factor count.

  • Add DQR_REQUEST label with a countdown for significant early-warning events.

  • Increase tukey window roll off to 1s to reduce biases due to windowing in bilby.

  • Increase request memory for online PE with IMRPhenomXPHM waveform model.

  • Enable parameter estimation on burst-BBH triggers associated with low-significance CBC triggers.

  • Update ROTA responsibilities in documentation.

  • Remove AGILE from the RAVEN workflow since it has been decommissioned.

  • Retry email bootstep when IMAP connection is reset.

  • Add gracedb client side 409 error code to the list of retries.

  • Don’t log omegascan FloatingPointErrors to Sentry since we know that they come from data which is only zeros (e.g., Virgo out of observing).

  • Filter Burst-BBH superevents from SNEWS coincidence searches.

  • Don’t apply EM_COINC if external event is SubGRBTargeted and vetoed by NOT_GRB.

  • Require no spaces and specific version format for filenames when sending update notices

  • Serialize changing of preferred external event.

  • Update Nagios URL and screenshot in the documentation.

2.3.6 “Champ” (2024-04-30)

  • Update trials factor to account for MLy joining O4b.

  • Update ligo-skymap to 2.0.0.

2.3.5 “Champ” (2024-04-11)

  • Remove oLIB from trials factor for beginning of O4b.

2.3.4 “Champ” (2024-04-04)

  • Fix bug that sent cWB-BBH notices as malformed Burst notices.

  • Change Bilby accounting group user back to to soichiro.morisaki.

2.3.3 “Champ” (2024-04-01)

  • Change Bilby accounting group user to cody.messick as temporary workaround while KAGRA accounts are in flux.

2.3.2 “Champ” (2024-03-22)

  • Enable public alerts on production.

2.3.1 “Champ” (2024-03-22)

  • Moved submit_rapidpe to inference.py and refactored to delete base_submit.

  • Add rates of external triggers as configuration variables. This allows us to make adjustments without a ligo-raven release.

  • Increase the number of parallel bilby jobs to increase the number of posterior samples.

  • Add search field to alert threshold and trials factor configuration variables.

  • Fix bug when creating combined sky maps in the targeted search, being loaded in the ring ordering rather than nested.

  • Update trials factor for O4b.

  • Pin bilby_pipe to bilby_pipe==1.3.1.

2.2.1 “Sheepsquatch” (2024-03-08)

  • Include p_astro and em_bright in alert payload for burst-cwb-bbh trigger.

  • Produce em_bright.json for burst-cwb-bbh triggers.

  • Add acceptance check to if dependencies have been approved by SCCB and if Bilby PE summary is available.

  • Do not add events involving KAGRA to superevents.

  • Require pesummary>=1.0.2 to fix bug in postprocessing of bilby results.

  • Update gwpy>=3.0.8 to fix issue with discontiguous lal caches.

  • Change the worker concurrency to 32 and set max-memory-per-child to 2GB.

  • Document librarian ROTA duties.

  • Make the dev configuration the default for testing.

  • Allow SSM events to be added to/create superevents. Disable em-bright on SSM events until implementation exists.

  • If an official Fermi sky map is already uploaded to external event, block future uploads.

  • Ignore events from VTInjection search with regards to annotations and superevent creation.

  • Disable sending alerts on production, leave them on for other instances.

  • Add a form to the Flask console to create BAYESTAR sky maps with selected detectors disabled.

  • Upgrade RapidPE/RIFT to v0.6.7, RapidPE to v0.1.1. Don’t set getenv = True in HTCondor, instead set getenv and environment options explicitly. Modify condor._submit, condor.base_submit, and condor.submit_rapidpe to accept an optional list for getenv, to pass to the condor_submit executable. Include pastro from RapidPE/RIFT in the second preliminary notice and initial notice+circular.

  • Add instruments and objid fields to the header of sky map files produced from posterior samples.

  • Ignore EarlyWarning triggers when launching RapidPE/RIFT.

  • Upgrade RapidPE/RIFT to v0.6.8.

2.1.10 “Bunyip” (12-05-2023)

  • Re-apply all v2.1.8 “Ahool” changes after emergency release of v2.1.9 “Ghost Deer”.

  • Prevent uploads of INTEGRAL Test events to the production server to avoid unnecessary polls of old GW data and use of resources.

  • Increase coverage for RAVEN test to nearly 100%; add notice type in gracedb log for external sky maps regarding their source.

  • Update documentation for RAVEN; add missing documentation for various other tasks.

  • Set a 600 second hard time limit for superevents.process tasks.

  • Set pipeline preferred events for low significance alerts. Previously, pipeline preferred events were only set for high significance alerts.

  • Check for stale caches by checking both the start-time and end-time of the segment. Previously only the start-time would be checked. Add Nagios check for existence of llhoft data.

  • Disable parameter estimation for offline G-events.

  • Add condor.submit_rapidpe, a submit task dedicated for RapidPE with lower retry_backoff_max than other PE pipelines.

  • Get the getenv value from the gwcelery.tasks.inference module variable instead of hardcoding in condor sub file.

  • Modify rapidpe config template to use specific options for high snr triggers. Turn-off cProfile for production and playground.

  • Update bilby to 2.2.2 and bilby_pipe to 1.3.0; ensure that environment variables are passed correctly when using HTCondor>=10.7.

  • Upgrade RapidPE/RIFT to v0.6.4, RapidPE to v0.1.1.

  • Update deployment to install gwcelery using pip instead of poetry because the latter installs the project in editable mode.

  • Update ligo-followup-advocate to 1.2.8.

2.1.9 “Ghost Deer” (11-27-2023)

  • Roll back all v2.1.8 “Ahool” changes for an emergency release of v2.1.7 “Spring-heeled Jack” with a bug fix for Out-of-Memory issues.

  • Require IGWN-Alert >= 0.3.0.

  • Add a separate em-bright worker for re-routing em-bright tasks.

2.1.8 “Ahool” (09-01-2023)

  • Sort Python packages on Flask webapp by name for readability.

  • Update Bilby to 2.1.2; fixes a post-processing bug that can impact distance posteriors for low-SNR events.

  • Update bilby_pipe to 1.2.0; address the same bug as for Bilby and fix a typo in frame paths on playground.

  • Update release template to refer and point to test instance rather than playground.

  • Modify project .bashrc to avoid sourcing a conda environment. Update deploy-common stage to install gwcelery using poetry.

  • Use the central time for querying events in the targeted subthreshold search. This is an improvement since the starting time of low SNR GRB candidates tends to be unreliable.

  • Use ligo-skymap API to create bayestar coherence plots.

2.1.7 “Spring-heeled Jack” (08-21-2023)

  • Specify GW group when triggering RAVEN pipeline if not given. This fixes a bug introduced when allowing the SubGRBTargeted search to find both CBC and Burst events.

  • Add condition that prevents a RAVEN alert if the FAR is negative. This is meant to add clarification that a SubGRBTargeted GRB with a FAR higher than the given threshold (which leads to a negative joint FAR), or any other situation that leads to a negative FAR does not pass publishing conditions.

  • Update ligo-followup-advocate to 1.2.6; enable option to remove line wrapping in order to conform to new GCN circular recommendations.

  • Update rapidpe-rift-pipe to 0.6.3. Rapidpe will run on triggers from all pipelines. In this version pTerrestrial is taken from the preferred event instead of calculating it from rankingstatpdf. Removed config options far_threshold and rankstat_pdf_file that were used for pTerrestrial calculation. Using config option total-points instead of points-per-side to set the required number of grid points in the initial intrinsic grid. total-points will make sure that we get close to the required number of grid points. Corrected the time-offset of MDC to match the current MDC. Setting the minimum frequency fmiin-template to 20Hz to match the minimum frequency used by bilby.

  • Fix typo to block coherence plots for combined sky maps, since these are redundant to their GW counterparts.

  • Add subthreshold targeted events from Fermi and Swift, as well as subthreshold untargeted events from Fermi to the O3 MDC replay.

2.1.6 “Champ” (08-10-2023)

  • Prevent Sentry from triggering on email notice validator since an error message is already logged to gracedb in case of failure.

  • Lower omegascan minimum frequency to 10 Hz from 20 Hz to better frequencies used by searches.

  • Change the superevent nagios check behavior to check against created time instead of t_0.

  • Install bleeding edge version of flower monitor until next point release fixing url-prefix option. For reference, see https://github.com/mher/flower/issues/1316

  • Don’t attempt to download em_bright and p_astro files from events that don’t provide them. This fixes the case where burst only alerts contained garbled information in the classification and properties section.

  • Expose PE skymap images, HTML, and gzip FITS when update notice is sent.

  • Prevent public tag being applied automatically to RapidPE_RIFT.p_astro.json. Therefore this is not included automatically at the time of sending initial alert.

2.1.5 “Ningen” (08-01-2023)

  • Add “online” label to PESummary result page from online Bilby.

  • Increase tolerance duration for superevent existence nagios check in the last hour from 1 min to 10 mins.

  • Add error handling to hop-client consumers, and restart the consumer on non-fatal errors.

  • Ensure that the content of coincidence_far.json is in JSON format in views.py.

2.1.4 “Mokele-Mbembe” (07-26-2023)

  • Use upload instead of create_log and update t_0 in update_if_dqok.

  • Add nagios check for presence of recent MDC superevents.

  • Add new celery worker for Kafka consumer.

  • For the targeted search, add configuration variables for FAR thresholds and additional trials factors. Prevent GRB candidates that fall outside of FAR thresholds from triggering an alert.

  • Update bilby_pipe to 1.1.2.

  • Updated RapidPE/RIFT Pipeline to 0.5.3, replacing errors from absent GstLAL triggers with a graceful failure. Updated RIFT to 0.0.15.9 to fix compatibility with newer NumPy versions. Updated LALSuite to 7.16 to fix bug in XLALSimInspiralTDModesFromPolarizations causing RIFT to fail. Setting appropriate prod/dev accounting group depending on GraceDB URL.

  • Disable online PE for MDC events from MBTA until multiple data sources can be handled.

  • Ensure consistent clim between omegascans and annotate plots with Q used. Add 2 seconds to start/end of data query to avoid whitening problems.

  • Change Kafka consumer configuration to listen to production Swift topic on all deployments and to listen to Fermi production topic on production deployment.

  • Disable automatic playground deployment; enable automatic test deployment.

  • Launch subthreshold targeted search for Burst events as well, using the same search windows.

  • Update ligo-followup-advocate to 1.2.5.

2.1.3 “Thunderbird” (2023-06-24)

  • Update to celery version 5.3.

  • Update ligo.em-bright to version 1.1.4.post2, ligo-followup-advocate to version 1.2.4.

  • Apply FROZEN_LABEL inside the handle_superevent as a part of the alert canvas so that they are retries for RetryableHTTPErrors.

  • Add ingestion of Fermi and Swift subthreshold targeted events. Generalize shared creation and replacement workflow used by various external trigger listeners into a single function.

  • Handle HTTP Error raised while grabbing Fermi skymaps.

  • Add handling for ValueErrors in detchar.check_vector to allow for frames without the requested channel.

  • Log failed omegascan exceptions to Sentry.

  • Change logging level of StateVector reading errors in detchar.check_vector back to log.exception now that https://github.com/gwpy/gwpy/issues/1211 is fixed.

2.1.2 “Ogopogo” (2023-06-09)

  • Update the low-significance threshold to be 2/day pre-trials factor. Update low-siginificance trials factors to 7.

  • Do not apply HIGH_PROFILE label to less-significant alerts.

  • Do not make pe skymap public automatically before approved.

  • Make image created when omegascan fails smaller.

  • Remove “overall state of detectors” line from detchar vector table, since without Virgo being online it will always be “bad”.

  • Check the iDQ OK vector before checking the value of iDQ timeseries.

  • Upgrade bilby_pipe to v1.1.1.

  • Enable SCiMMA kafka on minikube.

  • Use real KafkaError methods in delivery callback.

  • Use production accounting tag for bilby running on real events.

  • Disable querying data before starting PE as bilby now uses kafka data stream and PE timeout is long enough for low-latency frame data to be produced.

  • Move RAVEN tasks to main worker and make GCN ingestion sequential.

  • Update low-significance FAR threshold and trials factors to ensure these will go out at a rate of 2/day pre-trials factors.

  • Limit the number of MDC BBH bilby jobs to less than 5 to save disk usage.

  • Upgrade lalsuite to v7.15.

  • Updated rapidpe-rift-pipe version to 0.5.1. rapidpe will be submitted on superevents with gstlal/non-gstlal preferred event. Launching rapidpe 30s after merger. query_shm=True will run on lower latency data on /dev/shm/kafka/ if it is available, if not it will default to /ifocache/llcache/kafka/.

  • When updating the preferred event, also update t_0.

2.1.1 “Lone Island Mountain Devil” (2023-05-29)

  • Upgrade ligo-followup-advocate to v1.2.3.

  • Bump ligo.em-bright to v1.1.3 to be compatible with schema of online PE files with spinning waveforms.

  • Pin astropy to “!=5.3” until a bug plotting omega scans with gwpy is resolved.

2.1.0 “Jersey Devil” (2023-05-19)

  • Add GWSkyNet to annotate CBC all-sky search superevents. Only the superevents that meet the following conditions will be annotated by GWSkyNet: 1. A network of signal-to-noise ratio of greater than or equal to 7.0; 2. The SNRs in at least two detectors are no less than 4.5; 3. The FAR is below 1 / 21600Hz. In addition, superevents that meet the above conditions will also be annotated if there is a change of preferred event.

  • Only check iDQ for detectors that are on.

  • Update IGWN alert authentication documentation.

  • Clean up large bilby data files after the run for MDC injections.

  • Add pe tag to pesummary uploads.

  • Add new emfollow-dev deployment.

  • Ignore SSM search for superevents until decision about dedicated notice type and other details from OpsDiv are decided.

  • Update rapidpe-rift-pipe to v0.3.0. Adding config options to calculate pastro. Uploading rapidpe pastro to gracedb. Adding a link to rapidpe summary page in logs. Moving the config option to cprofile the jobs to the General section of config. Increasing the number of grid points per intrinsic coordinate to 10. Adding a config option to specify the type of run (online vs o3replay) when choosing channel names, the same as bilby_pipe does. Modified inference.py to set this config option based on the GraceDB URL.

  • Increase request memory of bilby runs for BBH.

  • Update bilby filenames in acceptance test.

  • Reduce GW FAR requirement for SOG_READY to 1/10 years.

  • Handle ValueErrors from checking iDQ channels, such as if the configured iDQ channel name is not the same as that in the frames. This happened for S230518h during ER15 and broke the alert pipeline.

  • Update the Kafka topic that we consume Fermi alerts from.

  • Fix the triple circular uploads by rearranging the alerts canvas and moving tasks into the second group.

  • Add ability to create circular for medium-latency GRB followup by PyGRB and X-pipeline from the dashboard.

  • Add pe tag to EM-bright probabilities calculated from bilby samples.

  • Add an automatic retry for the check_vectors task if requested data is not yet available in the caches due to data transfer latency. This will retry every five seconds up to four times, for a maximum total of 20 seconds.

  • Rerun the RAVEN sky map comparison whenever the GW preferred event changes.

  • Upgrade bilby to v2.1.1

  • Update ligo-raven to v3.2 and ligo-followup-advocate to v1.2.2.

2.0.6 “Spaghetti Tree” (2023-05-10)

  • Add HIGH_PROFILE label for rapid response team.

  • Ensure that external events are not already associated with a superevent when triggering Raven off of superevents.

  • Add minimum resolution for external sky maps.

  • Remove clutter from omegascan pngs and lower dpi to 150 to reduce file size.

  • Retry GCN circular creation upon 408 or 409 errors, likely due to AWS issues for GraceDB.

  • Add Kafka consumer bootstep.

  • Skip plotting sky map overlap integral if coinc_far_dict is empty, which occurs for SNEWS coincidences.

  • Update iDQ channel configuration for O4 iDQ channels. The playground config will continue to use the O3 channels to support the O3 MDC replay.

  • Update bilby settings. Set the chirp mass boundary between BNS and potential NSBH to 1.465Msun, computed from component masses of 3Msun and 1Msun, and use low-spin IMRPhenomD below and high-spin IMRPhenomPv2 above. For high mass triggers, use IMRPhenomXPHM. Use these “production” settings by default for MDC and production analyses, and we no longer have multiple bilby runs on a single event. Increase the time limit of condor job to 12 hours.

  • Bump pesummary version to 1.0.0. Disable ligo skymap and use 6 cpus and online-PE dedicated nodes to speed up the pesummary postprocess.

  • Enable public alerts in production.

  • Disable Kafka consumer bootstep in the production deployment.

  • Fix bug where Kafka alerts were using combined sky map from the external event rather than the superevent.

  • Fix bug where the external sky map was created with out-of-date external GCN information.

  • Require em-bright >= 1.1.2.

2.0.5 “Mothman” (2023-04-20)

  • Add Mattermost channel creation for the RRT to discuss a superevent candidate.

  • Use Online_PE_MDC nodes for bilby jobs on O3-Replay injections.

  • Do not remove PE run directory after the run completes or fails.

  • Filter external events based on External group in superevents.py, instead of logic based on substring matching.

  • Increase test coverage for views.py and fix manual mock coincidence injections.

  • Bump ligo.em-bright version to v1.1.1.

  • Bump ligo-followup-advocate version to v1.2.1.

  • Also update the preferred external event when issuing a manual RAVEN alert.

  • Skip less significant alert workflow in case of both significant and EW triggers. Previous workflow only accounted for blocking in presence of significant events.

  • Update rapidpe-rift-pipe version to 0.0.12. Use mchirp and q as the coordinates for the rectilinear intrinsic grid. Update accounting_group to ligo.dev.o4.cbc.pe.lalinferencerapid and add accouting_group_user. Add an option to map the events to an injection. Upload only a select set of log files to gracedb.

  • Prevent alternative collating method for O3 replay and MDC events with INTEGRAL.

  • Pick RAVEN_ALERT preferred external event over one that doesn’t pass the joint alert publishing criteria.

  • Avoid updating values in the combined sky map if missing in the GW sky map.

  • Catch bug if instrument not in external sky map header. This could occur if grabbing a sky map from an external URL such as from Fermi-GBM that lacks this field.

  • Prevent external event from switching superevents.

  • Only plot sky map overlap with prefered external event when RAVEN_ALERT label is applied.

  • Populate the duration and central_frequency fields in Kafka notices for burst events.

  • Add periodic SNEWS MDC events to test the corresponding IGWN alert listener.

  • Use sky map from preferred event rather than superevent, triggering off EM_READY label instead of SKYMAP_READY. If SKYMAP_READY is applied or if a sky map file is added to the superevent, we will once again try to get the GW sky map from the superevent.

  • Update bilby and bilby_pipe to 2.1.0 and 1.0.10 respectively. The number of spline nodes for calibration errors is increased to 10 thanks to the bilby optimizations. Change sampler settings into naccept=60, nlive=500.

  • Add automatic PESummary task to postprocess bilby parameter estimation results.

  • Indicate dispatch of first less-significant alert with LOW_SIGNIF_PRELIM_SENT. Launch timer for revision and send second less significant alert with the same label conditioned on blocking labels indicating significant alert or advocate action.

  • Replace EM_Selected and EM_SelectedConfident labels with LOW_SIGNIF_LOCKED and SIGNIF_LOCKED.

  • Fix repeating of raven alert publishing criteria met log message.

  • Don’t compute p-astro for spiir/AllSky because it now computes and uploads its own.

  • Prevent repeating of sky map comparison pipeline with the secondary flattened sky maps.

2.0.4 “Skunk Ape” (2023-03-28)

  • Add dashboard button to download sky map from a URL and upload to an external event.

  • Set pipeline preferred events after revising preferred event. Provide dashboard to set them.

  • Remove the prelimimary_alert_timeout variable and workflow, since the value is set to zero, and the pathway is no longer used.

  • Fix bugs in online bilby PE.

  • Use more production-like settings in online bilby PE for MDC triggers.

  • Clean up the superevent before sending significant alert.

  • Explicitly rank publishability of significant events higher than less-significant events ones in superevents.keyfunc since these are two distinct categories of alerts.

  • Remove infinite far threshold for earlywarning alerts for playground configuration. Set far threshold of 1/30 days for early warning triggers in conf/__init__.py.

  • Update the significant far threshold from 1/60 days to 1/30 days in conf/__init__.py.

  • Enable less-significant public alerts. Change the behavior for the EM_Selected label to lauch less-significant alert pipeline. Use the EM_SelectedConfident label to indicate completeness and passing significant full-bandwidth event; this starts the “significant” alert pipeline. Less-significant alert pipeline is blocked by the presence of EM_SelectedConfident.

  • Enable public early warning alerts. The superevent manager now applies the EARLY_WARNING label to the superevent when a significant EW event is added to the superevent. The automated pipeline is launched and is blocked before sending if EM_SelectedConfident is found to be applied.

  • Add O3 replay MDC testing with RAVEN pipeline. This will run on the emfollow-playground server, creating mock coincidences with a frequency given by the joint_O3_replay_freq variable.

  • Require sky map information to publish coincidence with a GRB candidate.

  • Fix collation of INTEGRAL notices by getting the GCN ID from the IVORN field.

  • Add criteria when choosing new preferred external event that real events are preferred over those likely not to be real, and SNEWS events won’t be overwritten by GRBs.

  • Update external coincidence to update circulars.

  • Relaunch RAVEN sky map comparison whenever a new GW or external sky map is available. Also copy over to the superevent if an alert has already been sent out.

  • Integrate MLy into superevent manager.

  • Update bilby to 2.0.0 and bilby_pipe to 1.0.8. Use ROQ bases dedicated for BNS mass region for low-mass signal. Switch to acceptance-walk method for dynesty sampling, and use naccept=10, nlive=500, and nparallel=2 for quick-BNS run, and naccept=20, nlive=1000, and nparallel=2 for the other runs.

  • Adjust arguments for ligo-skymap-from-samples: set the random seed to make the results reproducible and set the maximum number of points to limit run time.

  • Add condor accounting user tag to PE jobs.

  • Update rapidpe-rift-pipe to 0.0.8. This finds the initial grid region based on gstlal search biases. Fixes spin components to gstlal trigger spin. Offsets likelihood by 0.5*snr**2 to avoid numpy overflow. Reduced number of extrinsic samples and the number of extrinsic samples that get saved.

  • Pin redis<4.5.2 due to a potential bug discovered in redis 4.5.2.

  • Force omegascan plots to all use the same color scheme.

  • Plot longer timespan before event and shorter timespan after event in omegascans.

  • Take the minimum false alarm probability instead of the maximum when checking if iDQ false alarm probabilities pass the iDQ false-alarm-probability threshold.

  • Do not change current directory when setting up rapidpe dag.

  • Update condor memory request and accounting tag for sky map jobs.

  • Add a dedicated Celery worker for tasks that use Python multiprocessing for parallelism, separate from the queue for OpenMP parallelism. This should work around an issue where ligo-skymap-from-samples was using _both_ OpenMP and Python multiprocessing, running slowly and causing Condor job evictions due to excessive memory usage.

  • Add duration and central_frequency fields to Kafka notices. The fields are set to None for now.

  • Set the Significant field in VOEvent XML packets and Kafka notices.

  • Import Kafka notice schema from the new igwn-gwalert-schema library.

  • Drop support for Python 3.8.

2.0.3 “Ugly Merman” (2023-02-16)

  • Require bilby_pipe>=1.0.7.

  • Require matplotlib<3.7 to fix bleeding edge dependencies tests.

  • Use multi-order GW sky maps to produce combined sky maps and add combined sky maps to alerts alerts. The presence of the COMBINEDSKYMAP_READY label indicates the combined sky map is now available in that external event or superevent. Only copy a combined sky map to the superevent when sending an alert if the preferred external event has one available.

  • Resurrect unit test test_handle_superevent.

  • Make igwn alert listener retry all non-fatal errors. This should be revisited once https://github.com/astronomy-commons/adc-streaming/issues/54 is resolved.

  • Add EARLY_WARNING_LABEL symbol to superevents.py.

  • Remove subthreshold annotations from orchestrator canvases.

  • Fix race condition of multiple instances of the RAVEN pipeline by polling the superevent state before updating joint FAR.

  • Adjust PE event directory permissions.

  • Test SoG pipeline with MDC events.

  • Update ligo.em-bright to version 1.1.0.

  • Update BAYESTAR task to use PSD in coinc.xml file for all pipelines.

  • Convert cWB FITS files to multiorder format if not already in the multiorder format.

  • Don’t compute p-astro for PyCBC/AllSky because it now computes and uploads its own.

  • Don’t try to download psd.xml.gz in online PE, as it is no longer uploaded by any search pipelines.

  • Add functionality in superevent manager to decide confident vs. less significant alert criteria.

  • Re-purpose FROZEN_LABEL to mean subthreshold alert criteria is met. Add new SIGNIFICANT_LABEL to indicate presence of significant full-bandwidth trigger. Apply EARLY_WARNING_LABEL to superevent when there is a significant EW event.

2.0.2 “Flying Icarus” (2022-12-23)

  • Change web directory for PE outputs. Use Online PE nodes for bilby. Do not upload lalinference ini files.

  • Add RIFT analysis only for gstlal triggers.

  • Use the exttring worker for RAVEN tasks.

  • Bump ligo.em-bright to 1.1.0.dev1 to add HasMassGap. Review page: https://git.ligo.org/emfollow/em-properties/em-bright/-/wikis/Mass-gap-review

  • Compute source properties uniformly across all CBC pipelines.

  • Don’t compute p-astro for MBTA because they now compute and upload their own.

  • Change expected p-astro filename from p_astro.json to pipeline.p_astro.json, where pipeline is the name of the pipeline that uploaded the event.

  • Indicate a joint CBC-GRB event should be used for a measurement of the speed of gravity by applying the SOG_READY label to the superevent. This pipeline is launched is the right conditions are met after ADVOK label.

  • Add button to apply RAVEN alert labels to flask app. This will manually trigger a RAVEN alert.

  • Update messages from RAVEN alert pipeline to be more informative.

  • Fix bug where updated GRB events couldn’t create sky maps.

  • Filter out test GCNs from updating external events.

  • Update SNEWS ingestion to use canvas structure. This also fixes a bug where SNEWS tests events were not being ingested properly.

  • Fix datetime format in Avro and JSON notices so that they conform to ISO 8601. They were missing the T separating the date from the time and the final Z denoting the UTC time zone.

  • Set pipeline preferred events during superevent clean up.

2.0.1 “Alien Big Cat” (2022-10-26)

  • Restrict ligo.em_bright to >=1.0.4 to accept new posterior samples format

  • Fix regression in superevent clean-up

  • Fix regression where subthreshold annotations could use stale data

  • Silence Sentry for adc-streaming log messages

2.0.0 “Akkorokamui” (2022-10-14)

  • Introduce public alerts over Kafka via SCiMMA and the new GCN.

  • This release establishes the practice of naming GWCelery releases after cryptids.

  • Send SIGKILL to vacate condor jobs that do not die promptly when condor_rm’ed.

  • Preferentially pick coincident SNEWS events over coincident GRB events.

  • Add HTTP 408 Request Timeout errors to list of errors that trigger a retry.

  • Ignore BBH/IMBH specific searches for GW-GRB burst searches.

  • Perform first2years MDC uploads asynchronously.

  • Adjust request_memory specification for condor submission of OpenMP workers.

  • Ignore noisy adc-streaming log messages due to frequent but harmless errors in the IGWN Alert listener.

  • Replace instances of “MBTAOnline” with “MBTA”.

  • Remove gstlal from list of pipelines that upload PSDs in separate psd.xml.gz file. Update mock event uploads to include PSD in coinc.xml upload.

  • Disable LALInference parameter estimation and switch to Bilby as the main parameter estimation software.

  • Add bilby-pipe>=1.0.6 and gwdatafind>=1.1.1 dependency, and unpin pesummary to fix bilby workflow for O3-replay.

  • Disable parameter estimation for MDC events as it is currently broken for them.

  • Enable ingestion and processing of test SNEWS external events.

  • Run unit tests under Python 3.10.

  • Add DQR_REQUEST label to superevent after sending first preliminary alert.

  • Embed PSDs in first2years event uploads, matching the O4 configuration of most CBC pipelines.

  • Create external MDC events to test the RAVEN alert system. Test alerts include Fermi, Swift, INTEGRAL, and AGILE-MCAL. Add acceptance tests for the RAVEN alert system. Add options to use the superevent search field in coincident searches. Update the preferred external event based on the best available.

  • Report the URL of multi-resolution FITS files in GCN Notices and create flat-resolution files asynchronously.

  • Move functions in handle_grb_gcn to an asynchronous group to prevent detchar errors from interupting sky map generation.

  • Prevent sub-threhsold GRBs from overwriting high-threshold GRBs.

  • Listen to initial GBM alerts for earlier warning. Prevent these events from triggering alerts unless later updated.

  • Adjust Celery working concurrency settings: turn off autoscaling, turn on prefetching. This is seen to decrease the latency of most tasks. For example, the latency of gwcelery.tasks.gracedb.upload tasks is decreased by a median of about 0.5 s.

  • Set the preliminary alert timeout to 0 s.

  • Use multi-resolution GW sky maps when calculating the joint false alarm rate. Use single pixel RA/dec when evaluating for Swift coincidences.

  • Filter BBH/IMBH events from burst-GRB searches.

1.0.1 (2022-05-09)

  • Added request_disk specification for gwcelery condor submission

  • Bump p-astro to pre-release v1.0.0dev1. This version is a stop gap to bring back the reference O3 hard-cut implementation of p-astro, and resolve the dependency issues. Will need a full release later.

  • Add ligo.em_bright>=0.1.5 dependency.

  • Update MPICH module in deployment environment for RL8

  • Configure the playground environment to read O3ReplayMDC frames.

  • In O4, online CBC pipelines shall include the PSDs in the initial coinc.xml upload and shall not upload a separate psd.xml.gz file. Update the list of pipelines that have not yet made the transition and still require the old behavior (gstlal, spiir).

  • Set default RAVEN search to empty list to fix argument error.

  • Drop python3.7 support, add python3.9 testing.

  • Update to em-bright >=1.0 that implements EoS maginalization for HasNS and HasRemnant. An important difference compared to previous versions is that the trained classifiers are no longer stored as package data, but downloaded and cached using astropy.utils.data.download_file. These are also loaded in module scope in em_bright, and therefore we are no longer required to pass them explicitly.

  • Update conda environment in bashrc to igwn-py39-20220317.

  • Fix threshold to correct scale in order to consider a Fermi GRB real.

  • Prevent external GCN notices with no sky map information (all zeros) from creating sky maps.

  • Pass –no-deps to pip during deployment because of a bug in pip and because the poetry lock file already contains all of the dependencies

  • Drop –use-feature=in-tree-build from pip call in deployment, as pip now does this by default and has deprecated this option

  • Switch to using IGWN alerts instead of LVAlert. Add igwn-alert as a dependency.

1.0.0 (2022-01-21)

  • Update to Celery 5.

  • Rename branch master to main.

  • Switch build, packaging, and deployment from setuptools+pipenv to poetry.

  • Use a date-tagged IGWN Conda environment to prevent unversioned changes to dependencies.

  • Some unit tests now use a live worker instead of “eager” mode. As a result, Celery’s behavior in those unit tests is more similar to production, and therefore more likely to catch any concurrency bugs, race conditions, or deadlocks.

  • Rewrite GitLab CI pipeline to use the IGWN computing group’s Python job templates.

  • Use ssh+kerberos instead of the now-defunct gsissh for unattended login to LDG hosts for deployment jobs in the GitLab CI pipeline.

  • Remove mock module imports from Sphinx configuration, for simpler and more robust documentation builds.

  • Increase the value of the Celery worker_proc_alive_timeout to 8 seconds in order to avoid unnecessarily killing workers that are slow to start up.

  • Remove workarounds that were in place to preserve order of results from groups of tasks, because Celery 5 now preserves result order automatically.

  • Require astropy >= 4.3.1 due to an upstream bug (https://github.com/astropy/astropy/issues/11879).

  • Fix a bug in configuration of the Jinja template directory.

  • Reduce queries to gracedb by RAVEN by passing event dictionaries directly.

  • Switch to use gracedb-sdk for RAVEN.

0.13.1 (2021-03-01)

This release primarily updates versions of dependencies.

0.13.0 (2020-06-03)

  • Set FAR threshold for early warning alerts to once per day.

  • Identify early-warning events using the EARLY_WARNING label rather than the EarlyWarning search type. The search type is already used to distinguish mock (MDC) events, so it cannot also be used to indicate early-warning events.

  • Inhibit GCNs for superevents with the INJ label.

  • Add configuration variable to disable all but MDC alerts from GCN, and set that variable to True on the production instance.

  • Skip the preliminary alert timeout for early warning events.

  • Update the documentation on RAVEN functions and external triggers flow chart.

  • Change BAYESTAR low frequency cutoff from 30 Hz (the default value) to 15 Hz.

  • Change playground configuration to read O3 replay data rather than O2 replay data.

  • Drop dependency on seaborn.

  • Defer sleekxmpp imports until the VOEvent client starts. This way, sleekxmpp is only imported in the thread that actually uses it. This should speed up worker startup by about 0.1 seconds.

  • Defer Comet and Twisted imports until they are actually needed by the VOEvent broker. This should speed up worker startup by about 0.2 seconds.

  • Defer imapclient imports until the email client starts. This should speed up worker startup by about 0.1 seconds.

  • Improve robustness of the email connection by resetting IMAP IDLE mode at least every 5 minutes and improving error-checking upon disconnection.

  • Add platform and hostname information to the Flask dashboard.

  • Ensure external sky maps are normalized.

0.12.3 (2020-03-24)

  • Fix a bug that was introduced in GWCelery 0.12.1 that broke BAYESTAR localizations for PyCBC events. In 0.12.1, the BAYESTAR data handling was changed to merge together the contents of the coinc.xml and psd.xml.gz files into a single XML document so that BAYESTAR was not sensitive to the order in which the two files were passed to it. PyCBC includes the PSD data in its initial upload, and so its psd.xml.gz file is just a copy of coinc.xml. Merging the two documents together resulted in a single file with two copies of every LIGO-LW table, which broke subsequent parsing.

    Fix this by adding a special case for PyCBC to download the coinc.xml file only. This has the nice side effect of reducing the latency for PyCBC events because it is no longer necessary to wait for the additional GraceDB REST API calls involved in uploading and download the additional file.

0.12.2 (2020-03-20)

  • Skip detchar checks for events which occur in the future.

  • Delay omegascans until data are available for events in the future.

  • Enable Zstandard compression of tasks and results to reduce bandwidth into and out of Redis.

  • Enable receipt confirmation of early warning GCN notices.

  • If available, use spatial coincidence FAR to determine when to publish a coincident event. Update both time and spatial FAR within superevent when publishable.

  • Fix bug where the superevent handler can trigger on external events.

0.12.1 (2020-03-12)

  • Set delay to produce preliminary alert to 0 seconds in the playground configuration. In the production configuration, the delay is still 30 seconds.

  • Adjust broker transport and worker settings so that the superevent worker respects task priorities. This is seen to reduce the latency of preliminary alerts by about 10 seconds.

  • The gwcelery.tasks.bayestar.localize task no longer cares about the order in which the coinc.xml and psd.xml.gz file contents arguments are passed to it because the task now combines the XML documents using ligolw_add. This allows us to change the immediately upstream task in the localization canvas from an ordered_group to a group. This avoids extra trips of the large file contents blobs into and out of Redis. This is seen to reduce the latency of the localization by about a second.

  • Produce GCN notices of type LVC_EARLY_WARNING for events that have the EarlyWarning search tag.

  • Add a new configuration variable early_warning_alert_far_threshold to control the FAR threshold for early warning alerts. In the playground environment, its value is the same as the threshold for ordinary CBC events. In the playground environment, its value is infinity, to generate alerts for all early warning events.

  • Fix bug where a SubGRBTargeted event would trigger a search in both Fermi and Swift.

0.12.0 (2020-03-05)

  • Add the GCN_PRELIM_SENT label after the GCN notice has been sent. Previously, the GCN_PRELIM_SENT label was added after the GCN notice had been sent and after the GCN Circular template had been created. Since it takes many tens of seconds to create the GCN Circular template, this was distorting latency statistics.

  • Prioritize processing of label_added LVAlert messages over new LVAlert messages in the superevent manager. The labels SKYMAP_READY, EMBRIGHT_READY, and PASTRO_READY must all be present before we can send a public alert, so processing label_added messages with higher priority may speed up preliminary alerts.

  • Increase the minimum concurrency of the main GWCelery worker pool from 4 to 8 subprocesses in order to decrease latency.

  • Append to, and do not overwrite, log files, when starting GWCelery via Condor.

  • Launch raven coincidence search for sub-threshold GRBs separately for different gamma-ray experiments in order to use different time windows. This enables the joint LVK-Fermi and LVK-Swift targeted searches to be integrated with RAVEN.

  • Grab subGRB Fermi sky maps from GCN.

  • Create external sky maps for offline subGRBTargeted Swift uploads.

0.11.2 (2020-02-28)

  • Document recommended value for the Redis server setting client-output-buffer-limit in order to prevent disconnection of Celery workers returning large task results. This value was established early in O3, but since it was not in the documentation, we frequently forgot to set it when configuring a Redis server on a new or upgraded system.

  • Add the unit test for tasks/inference.py.

  • Upload LALInference DAG files to save the exact commands run for the parameter estimation.

  • Fix the file names of Bayeswave PSDs.

  • Capture an exception that is produced when attempting to make an omega scan of data that contains NaNs.

  • Catch missing trigger_duration when launching check vectors for external events.

  • Run unit tests for Python 3.8.

  • Update ligo-followup-advocate to 1.1.6.

  • Update gracedb-sdk to 0.1.4.

  • When a GRB or SNEWS GCN is received, upload it to GraceDB with the correct group depending on the value of the VOEvent role attribute: if role="test", then upload to the Test group; if role="observation", then upload to the External group.

0.11.1 (2020-02-21)

  • Un-pin LALSuite and use the latest stable version (at this time, 6.68).

  • Do not use Online_PE condor slots for lalinference parameter estimation.

0.11.0 (2020-02-21)

  • Use Online_PE condor slots for lalinference parameter estimation.

  • Use Bayeswave PSD for online PE.

  • Fix a bug in skymap generation with online PE posterior samples.

  • Reduce the number of bilby runs for test events to less than once per day.

  • Add systematic error contributions to Fermi-GBM sky maps.

  • Convert Swift-BAT error radii from 90% C.L. to 1-sigma.

  • Add INTEGRAL and AGILE MCAL to GRB pipelines.

  • Apply label NOT_GRB to external Fermi candidates that do not meet required threshold of a GRB. This is determined by Most_Likely_Index and Most_Likely_Prob quantities supplied with Fermi notices. RAVEN will not consider external events labeled NOT_GRB.

  • Automatically generate and upload a graphic showing the source property values by means of a bar chart.

  • Pin astropy < 4.0 to work around an issue with caching of downloaded data on the Caltech cluster. See https://github.com/astropy/astropy/issues/9970.

  • Switch from GraceDB REST API calls from gracedb-client to gracedb-sdk to gain increased transaction throughput due to HTTP connection pooling.

  • Remove vetted=True keyword argument for GraceDB API calls to produce VOEvents, because that argument was removed from the GraceDB server and client over a year ago.

0.10.0 (2020-02-07)

  • Decrease the number of OpenMP workers from 40 to 16, now that gstlal is uploading a reduced number of events.

  • Add VOEvent broker and receiver configuration for playground environment in order to enable end-to-end testing of transmission to and receipt from GCN.

  • Fix a bug in the upload of bilby results.

  • Do not start parameter estimation for mock events uploaded to gracedb.ligo.org.

  • Calculate joint spatio-temporal FAR automatically for external coincidences. Create the combined skymap when both the GW and external skymaps are available.

  • Increase the number of retries, with incremental retry backoff, when fetching the Fermi sky map from HEASARC. This is because the Fermi skymap is typically uploaded tens of minutes after the GCN notice from Fermi.

0.9.2 (2020-01-07)

  • Update to Celery 4.4.0.

  • Add bullet charts for BAYESTAR coherence-versus-incoherence Bayes factors. The BAYESTAR log Bayes factor for coherence versus incoherence is stored in the FITS file header’s LOGBCI field. For each FITS file that has this header field, make a bullet chart to compare the log Bayes factor to a standard table of threshold confidence levels from Kass & Raftery (1995).

  • Enable the RAVEN alert pipeline by having the superevent manager listen to the label RAVEN_ALERT.

  • Use RAVEN VOEvent if RAVEN_ALERT.

  • Generate emcoinc circular if RAVEN_ALERT instead of EM_COINC.

  • Increase both CBC and Burst trials factors by one due to enabling the RAVEN pipeline.

  • Refactor gwcelery.tasks.detchar.make_omegascan to reuse GWPy’s own plotting functions, instead of using our own Matplotlib code. This fixes a bug that prevented make_omegascan from working with Astropy 4.0 or later.

  • Unpin Astropy version, now that make_omegascan works with the most recent version.

0.9.1 (2019-12-15)

  • Produce an ADVREQ notification as soon as there is an alert which meets the public alert threshold, regardless of whether its annotations are complete. As a result, follow-up advocates will usually receive notifications about 30 seconds earlier, and will receive notifications even if some of the annotations fail.

  • Increase the FAR threshold of online PE to the public alert threshold.

  • Update lalsuite to lalsuite==6.63 and unpin scipy.

  • Change RAVEN to grab sky map from superevent. Block joint FAR calculation for SNEWS coincidences.

  • Skip Virgo data when online PE is started on O2Replay data since its statevector cannot be read by gwpy.

  • Modify RAVEN to run on MDC events.

  • Restrict the superevents.process task to process only complete G events instead of running for all the superevent completeness labels. The behavior for running on the new type events remains unchanged.

0.9.0 (2019-11-23)

This is the initial release of GWCelery for O3b.

  • Changes related to configuration settings

    • Use the Redis server that is provided by the operating system (e.g. as a systemd unit) rather than starting our own Redis server. This prevents a race condition between the shutdown of Redis and the shutdown of the workers that caused the workers to hang on shutdown.

    • Update HTCondor accounting group from O2 to O3.

    • Increase throughput for sky localization tasks by offloading processing of the openmp Celery queue to 40 workers that are launched via HTCondor on specially configured cluster nodes.

    • Use mpich as the MPI runtime for parameter estimation.

    • Use different HTCondor accounting groups for Celery workers depending on whether GWCelery is running in the playground environment (ligo.dev.o3.cbc.pe.bayestar) or the production environment (ligo.prod.o3.cbc.pe.bayestar).

    • Drop support for Python 3.6 so that we can use the check_output keyword argument that was added to suprocess.run() in Python 3.7.

    • Pin gwpy to <= 0.15.0 since the updated gwpy fails to read Virgo’s state vector.

    • Update ligo-followup-advocate version to 1.1.3.

  • Changes related to superevent/orchestrator design

    • Add event completeness to publishability criterion. All three of PASTRO_READY, SKYMAP_READY, and EMBRIGHT_READY will be used to evaluate event completeness for CBC events. Only the SKYMAP_READY label will be used to evaluate completeness for burst events.

    • Use EM_Selected to freeze the preferred event of a superevent and launch a preliminary alert.

    • Make sub-threshold annotations independent of annotations for superevents which pass public alert threshold.

    • Prevent second preliminary to be sent in the event of any advocate action. Previously, this was only being prevented for ADVNO.

    • Make skymaps from parameter estimation public automatically.

  • Changes related to online parameter estimation

    • Move a comment attached to posterior samples to the corresponding skymap.

    • Add bilby online parameter inference workflow.

    • Fix approximant name used for automatic parameter estimation.

    • Start parameter estimation on mock events.

    • Add acceptance tests of parameter estimation.

    • Use nodes dedicated to online PE also for playground events so that the test runs do not get stuck due to the lack of resources.

    • Add spins in online PE on playground events so that embright probabilities are calculated based on the posterior samples without errors.

    • Remove skymap generation from PE DAG so that it will not be generated twice.

    • Notify which pe pipeline failed for the failure of pe condor jobs.

  • Changes related to external coincidences

    • Create RAVEN circular if EM_COINC label is applied to superevent.

    • Make coincidence FAR synchronous within RAVEN pipeline to fix race condition.

    • Remove redundant SNEWS handler key.

    • Remove generation of em_coinc circular when EM_COINC label is applied.

    • Apply EM_COINC to preferred event when coincidence passes RAVEN publishing conditions.

    • Attempt fetching and uploading Fermi skymap upon receinving GCN notice.

  • Changes related to skymap generation

    • Revert back to running BAYESTAR for all G events.

    • Pass the -j flag to ligo-skymap-from-samples to speed up skymap generation.

  • Changes related to automated data quality checks

    • Create omegascans for all detectors upon creation of new superevent.

    • Run check_vectors upon the creation of a superevent. This will allow subthreshold superevents to be annotated with DQOK or DQV label.

  • Changes to the Flask dashboard

    • Teach preliminary alert form in Flask dashboard to present a dropdown of events sorted by the preferred event criterion.

    • Incorporate update circular into flask app.

  • Other changes

    • Add a task to em_bright.py to compute and upload source properties upon the upload of LALInference.posterior_samples.hdf5.

0.8.7 (2019-09-14)

  • Update ligo-raven version to 1.17.

0.8.5.1 (2019-09-04)

This is a non-sequential bugfix release based on version 0.8.5.1 to fix the following issue:

  • Fix a lethal bug in em_bright.py introduced in version 0.8.5. The bug would incorrectly use the snr as the maximum mass of the NS and therefore the source property estimation for pipelines apart from gstlal would be grossly incorrect.

0.8.6 (2019-09-01)

  • Update ligo-raven version to 1.16.

  • Fix a bug that prevented retrying of failed GraceDB API calls in the superevent manager.

  • Add a retry for one more potential GraceDB API failure in the initial and update alert workflows.

  • In the playground environment only, upload each mock event several times in rapid succession with random jitter in order to simulate multiple pipeline uploads.

  • Expose events to the public prior to sending any kind of alert: preliminary, initial, update, or retraction. Previously this behavior only occurred for preliminary alerts, which created the unusual and undesirable possibility of a public GCN for an event that is not public. As before, events are only exposed to the public in the production environment, and not in the playground environment.

  • Propagate sky map file extensions (as in bayestar.fits.gz,1) to the URLs that are presented in GCN notices.

  • Generate flattened FITS files and sky map visualizations for all superevents, even those that do not rise to the public alert threshold. Note that as a side effect all superevents will have the EM_Selected label applied, since it is used as a semaphore to trigger the annotations. The ADVREQ label used to serve double duty as the semaphore and also as the wake-up call for follow-up advocates, but now it only serves the latter purpose.

    The feature of generating flattened FITS files and sky map plots for all superevents comes as a request from the Fermi and Swift sub-threshold searches.

  • Delay running BAYESTAR until the superevent’s preferred event has stabilized. BAYESTAR is the most computationally intensive postprocessing task and running it for all events belonging to a superevent was a bottleneck.

  • For the playground environment only, decrease the timeout for stabilization of the preferred event from 5 minutes to 2 minutes, which is comparable to how long it has taken recent events to settle. This does not affect the configuration of the production environment.

  • Changed handle_cbc_event handler to call gstlal trained ML based inference for source property computation for gstlal triggers.

  • Apply EM_COINC to superevent and external event in parallel.

0.8.5 (2019-08-23)

  • Made raven.py tests more robust and have increased coverage.

  • Removed the feature of p-astro and em_bright reading mean counts, livetimes or ML classifiers from emfollow/data; moved them to lscsoft/p_astro as package data. Added back the test_tasks_p_astro.py that was accidentally taken out in v0.8.0. Pinned p_astro == 0.8.0.

0.8.4 (2019-08-16)

  • Fixed a bug in gracedb.create_tag to handle the scenario when multiple log messages exist for the same filename. The tag is applied to the most recent log message.

  • Retry GraceDB API calls that fail due to receiving incomplete or malformed HTTP responses, as indicated by http.client.HTTPException exceptions. This should work around the increased incidence of RemoteDisconnected exceptions that we have seen recently, and that caused a delay in sending out the preliminary alert for S190814bv.

0.8.3 (2019-08-09)

  • Enable Redis integration for Sentry error reporting.

  • Downgrade lalsuite to 6.54 since lalinference_pipe in 6.59 has a minor bug, which breaks automatic parameter estimation.

  • Include the number of participating detectors in the preferred event selection criterion for compact binaries: 3-detector events should be preferred over 2-detector events, and 2-detector events over 1-detector events, on the basis of more accurate localization. Ties are still broken by SNR.

  • Catch SystemExit exceptions from Python command line tools called in Celery tasks and re-raise them as RuntimeError exceptions so that they do not cause the workers to exit.

0.8.2 (2019-08-02)

  • Apply the public tag to data products before sending out an update GCN notice. This will prevent human errors related to not exposing LALInference files before sending a GCN notice.

  • Don’t read the entire HTTP response from GraceDB POST requests. We only need the HTTP status code. This change might speed up GraceDB API calls a little bit.

  • Increase preliminary alert timeout back to 5 minutes.

  • Make gracedb.create_superevent, gracedb.update_superevent and gracedb.add_event_to_superevent idempotent by catching the HTTPError from GraceDB that occurs if the superevent has already been created.

  • Fix bug where neither the space-time nor temporal coincidence far is calculated if external sky map is unavailable.

  • Update ligo.skymap to 0.1.9. This version changes the data type of the multi-resolution HEALPix format’s UNIQ column from an unsigned integer to a signed integer.

    Starting with this version, the Linux builds of ligo.skymap are compiled and optimized using the Intel C Compiler.

  • Change the trials factor for CBC searches to 4, since SPIIR is performing a single search, and that for burst to 3, since oLIB is not currently in operation.

0.8.1 (2019-07-29)

  • Downgrade lalsuite to 6.59.

  • Revert change that tried to fix incorrect key for querying external events. The keys were correct before.

0.8.0 (2019-07-26)

  • Assign gwcelery.tasks.skymaps.plot_volume tasks a reduced Celery priority as compared to gwcelery.tasks.bayestar.localize so that the latter are given preference. This ought to speed up the preparation of preliminary GCN notices because only the latter are required for GCNs but both kinds of tasks compete for slots in the resource-intensive OpenMP queue.

  • Reduce priority for CBC annotation tasks for events that do not pass the public alert threshold.

  • Update lalsuite to 6.60.

  • Ensure gracedb calls to create and update superevents are retried in the event of transient GraceDB API errors.

  • Update ligo-raven version to 1.15. Apply EM_COINC label in raven.py to give more control and prevent race conditions.

  • Use the space-time coincidence FAR as the default for RAVEN, use the temporal coincidence FAR when sky maps are not available.

  • Check if GRB is sub-threshold, set search to be ‘SubGRB’. Pass search through external triggers pipeline and RAVEN.

  • Tune Celery’s result_expires setting from its default value of one day to five minutes. Since we pass large byte strings as task arguments and return values, one day is too long to keep task tombstones in the database. This adjustment should reduce the memory footprint of the Redis server during periods with very high rates of GraceDB uploads.

    The downside is that task details will remain browsable in Flower for a much shorter period.

  • Remove p_astro_gstlal.py module, corresponding test modules, and documentation; p_astro will be reported as a pipeline product from gstlal. The computation of p_astro for all other pipelines is unaffected.

  • Fix EM_COINC bug where it is being over-applied to superevents.

  • Fix bug where wrong key was called for querying external events.

0.7.1 (2019-07-12)

  • The initial alert workflow will now consider only *.fits.gz sky maps and not *.fits sky maps for GCN Notices. It was an oversight that we did not exclude *.fits files from the list of extensions to consider when we updated the handling of multiresolution sky maps.

  • Catch and retry HTTP 429 (“Too Many Requests”) errors from GraceDB.

  • Enable Sentry integration for Tornado in order to capture errors from the Flower console.

  • Fix file extensions for LALInference sky map PNG files: they should be named LALInference.png, not LALInference.multiorder.png.

  • Increase the Redis server’s log verbosity in order to help diagnose Redis client connection dropouts.

  • Run sky map plotting and annotation tasks asynchronously so that they do not block sending preliminary alerts. Their outputs are only for human consumption; they are not needed in order to prepare GCN Notices.

0.7.0 (2019-06-21)

  • Trigger a preliminary alert for a superevent upon the first time that the preferred event is set to an event that meets the public alert criterion.

    This fixes a longstanding issue that has prevented automated preliminary alerts from being sent so far. The preferred event at the instant that the timeout ended did not meet the public alert criterion, but a preferred event that was selected some tens of seconds later did.

  • Decrease preliminary alert timeout to one minute.

  • The combined effect of these changes should be to decrease the latency for producing preliminary alerts from 7 minutes to 2 minutes.

0.6.3 (2019-06-14)

  • Work around a Celery canvas bug that prevented LALInference postprocessing from completing.

  • Fix a copy-paste error that caused DQV and INJ labels to be ignored when determining whether to send a preliminary alert.

  • Move RAVEN time coincidence windows to the application configuration.

  • Document the acceptence tests checklist in the instructions for preparing a release.

  • Update ligo-raven to version 1.14.

0.6.2 (2019-06-07)

  • Add a dependency on dnspython to silence the following warning message from SleekXMPP:

    DNS: dnspython not found. Can not use SRV lookup.
    
  • Pin some recently updated dependencies of Celery that caused unit test failures: amqp <= 2.4.2, kombu <= 4.5.0, vine <= 1.3.0.

  • Prevent subthreshold GRBs with low reliability from being processed as external events.

  • Add a task in orchestrator.py to generate FITS files and sky map images automatically whenever an HDF5 posterior samples file is uploaded.

  • Remove special-case handling of single-instrument events. Now, the eligibility of an event for a public alert is determined only on the basis of its false alarm rate.

  • Run parameter estimation on nodes dedicted to online-PE.

  • Emcoinc circular is triggered when RAVEN uploads a coincident FAR.

  • Pin scipy since scipy>=1.3.0 removes an interpolation function which lalinference postprocessing requires.

0.6.1 (2019-05-24)

  • Work around a bug in the Sentry Python SDK that caused excessive reporting of certain GraceDB exceptions that are listed in tasks’ autoretry_for settings. See getsentry/sentry-python#370.

  • Change the name of BAYESTAR localization files to bayestar.multiorder.fits to distinguish them from flat-resolution HEALPix files, which are still named bayestar.fits.gz.

  • Reimplement LVAlert listener as a Celery bootstep to avoid needing to track a singleton task using a Redis lock, because Redis locks do not play nicely with Redis persistence. The --lvalert command line option must now be passed in order to enable the LVAlert listener.

  • Turn on Redis database persistence so that Celery task state is preserved across restarts.

  • Add expose_to_public setting to disable exposing GraceDB events to the public in all environments except for production.

  • Update to the latest version of GWPy and un-pin Matplotlib because GWPy now supports Matplotlib 3.1.

  • Pin LALSuite to version 6.54 because LALInference in LALSuite 6.55 is not compatible with Python 3.

0.6.0 (2019-05-20)

  • Work around a bug in complex Celery canvases (see celery/celery#5512) that prevented initial GCN notices from being sent. As a side effect of this workaround, the initial, update, and retraction canvases will not automatically expose events to the public.

    The preliminary alert canvas still does expose events to the public, so under normal circumstances, the follow-up advocate should not have to manually do that. However, if the event has not been exposed to the public for whatever reason, then the follow-up advocate should expose it to the public manually before applying the ADVOK label. See emfollow/followup-advocate-guide!2.

  • Reduce the false alarm rate threshold for parameter estimation to decrease cluster load.

  • Remove redundant LVAlert subscription in handle_lvalert_grb to prevent double calls to RAVEN.

  • Read template weights for P_astro from hdf5 file using h5py for speedup.

  • Require matplotlib < 3.1 becuase matplotlib 3.1 breaks importing gwpy:

    /usr/local/lib/python3.7/site-packages/gwpy/plot/rc.py:79: in <module>
        rcParams.get('text.latex.preamble', []) + tex.MACROS),
    E   TypeError: can only concatenate str (not "list") to str
    
  • Make gwcelery.tasks.gracedb.get_superevents and gwcelery.tasks.gracedb.get_events take any number of keyword arguments to be passed to corresponding client methods.

  • Update the superevent t_0 field whenever the preferred event changes.

0.5.7 (2019-05-13)

  • If the VOEvent broker is disabled by setting voevent_broker_whitelist to an empty list, then suppress the normal error message that would occur when attempting to send a VOEvent when there are no broker connections.

  • Rearrange preliminary alert workflow so that sky map plots are generated for the newly added FITS file rather than an older FITS file that coincidentally has the same name.

  • Have gwcelery.detchar.check_vectors task apply all GraceDB log messages in order to increase robustness to recoverable GraceDB API errors.

  • Port over majority of P_astro code from gwcelery to the p-astro package.

  • Use cleaned data for parameter estimation.

  • The DQOK and DQV labels should be mutually exclusive. When gwcelery.tasks.detchar.check_vectors adds one of the DQOK or DQV labels, it will now first remove the other label.

  • Change exception in VOEevent parsing of Fermi subtreshold alerts to match real incoming alerts.

  • Update Celery to 4.3.0.

  • Automatically select the most up-to-date calibration uncertainties for parameter estimation.

0.5.6 (2018-05-08)

  • Extend the orchestrator_timeout to 300s and the pe_timeout to 345s. The previous timeout was not sufficient for the online pipelines to upload all of their possible candidates, hence the extension.

0.5.5 (2019-05-03)

  • Cycle through llhoft, high latency frames, and low latency frames in detchar’s cache creation.

  • Add explanations on options in online_pe.jinja2 for those who start parameter estimation based on the ini files uploaded to GraceDB.

  • Calculate horizon distance with psd.xml.gz to determine the upper limit of distance prior for parameter estimation.

  • Start parameter estimation when the lowest FAR of the events in a superevent is lower than the threshold.

  • Update the calibration uncertainties used for parameter estimation.

  • Handle an exception in VOEvent parsing of Fermi subthreshold alerts due to different param names.

  • Stop uploading corner plots of intrinsic parameters.

  • Connect to different GCN servers to receive alerts in the production and playground environments, because GCN does not support multiple receiver connections from the same client IP address to the same server.

  • Change the preferred event assignment logic to not let accidental candidates like G330298 which have low FAR but high SNR values to become the preferred event. From now on, superevents.should_publish takes maximum precedence for selecting the preferred event. The same is also used by orchestrator to expose events.

  • Update RAVEN coinc FAR task call which uses string params versus un-pickleable class object params.

  • Make sure to consume the entire response from every GraceDB API request. This will ensure that GraceDB API call has completed before the pipeline continues, and will fix errors like we encountered with S190426c where the pipeline would march along before uploads had finished.

  • Apply ADVREQ label earlier in the preliminary alert workflow.

  • Update LALSuite to version 6.54. We are now using a stable version again instead of a nightly build.

  • Add Nagios checks for GCN connectivity.

  • Improve uploaded comments so that it is easily understood which event has triggered parameter estimation.

0.5.4 (2019-05-01)

  • Provide a value for terrestrial count for P_astro for non-gstlal pipelines that is consistent with the FAR threshold used.

0.5.3 (2019-04-17)

  • Update ligo-followup-advocate to 0.0.28.

  • Stop using unreviewed cleaned data for parameter estimation.

  • Update detchar check to analyze full template duration for CBC events.

0.5.2 (2019-04-15)

  • Fix typo in gracedb.get_instruments: there was the attribute lookup single.ifo, which should have been the dictionary lookup single[ifo].

  • Fix gwcelery.tasks.p_astro_other.choose_snr for gstlal. This method did not previously expect to be called for gstlal, since it is typically only invoked for other pipelines. However, there is one case when choose_snr is invoked for gstlal, which is when the ranking_data file from gstlal is corrupted with NaNs, causing P_astro for gstlal to fail. Thus, choose_snr has now been fixed to also handle gstlal as a pipeline.

0.5.1 (2019-04-12)

  • Changed default for em-bright from 2.83 to 3.0 M_sun to be consistent with notices.

0.5.0 (2019-04-12)

  • Give permissions to read the files under parameter estimation run directories to non-owner people so that rota people can check their progresses. The naming convention of the run directories changed.

  • EM-Bright ML classification requires review. Until then, give answer based on low-latency estimates.

  • Compute P_astro with mass-based template weighting. Template weights are now keyed on template parameters, rather than bin numbers. This should make P_astro immune to binning conventions.

  • Add form to manually send a preliminary GCN Notice.

  • Fix a typo in gwcelery.sub that caused the Flower dashboard to fail to start.

  • Round iDQ p(glitch) to 3 decimal places in GraceDB log message.

  • Switch log telemetry from the on-premise instance of Sentry at Caltech to a cloud-hosted subscription to sentry.io.

  • In the playground configuration, the gwcelery.tasks.gcn.validate task was producing false alarms because the GCN receiver was receiving VOEvents from the production instance, which would certainly differ in content from VOEvents in the playground instance. Fix this by having gwcelery.tasks.gcn.validate discard all VOEvents if the VOEvent broadcaster is disabled.

  • Update ligo-followup-advocate to 0.0.27.

  • Wait for 1 minute before parameter estimation in case the preferred event is updated with high latency.

  • Ensure that P_astro accounts for very loud MBTA and PyCBC events, whose FAR saturate at certain low values depending on instrument combination, but whose SNRs can increase indefinitely.

  • When a user triggers a Preliminary or Update alert through the Flask interface, create a GraceDB log message to record the username.

  • The Flask interface will now show a confirmation dialog before sending any alerts.

  • Add a terrifying warning to the Flask interface to make it clear that the interface is live.

0.4.3 (2019-04-05)

0.4.2 (2019-04-05)

  • Calculation of number of instruments is now unified across superevent manager and orchestrator using gracedb method get_number_of_instruments.

  • Enable automated preliminary alerts for all pipelines because disabling them in the orchestrator introduced some issues due to the criteria for releasing a public alert drifting away from the definition of a the preferred event of a superevent. We will instead trust pipelines that are still under review will upload events to the playground rather than the production environment.

0.4.1 (2019-04-02)

  • Fixed normalization issues with p_astro_gstlal.py; normalization was being applied in the wrong places during Bayes factor computation.

  • Require celery < 4.3.0 because that version breaks the nagios unit tests.

  • Update false alarm rate trials factors for preliminary alerts.

  • Enable sending GCN notices for fully automated preliminary alerts.

  • Add threshold_snr option in online_pe.jinja2, which is used to determine the upper limit of distance prior.

  • Use the same criteria to decide whether to expose an event publicly in GraceDB as we use to decide whether to issue a public alert.

  • Do not issue public alerts for single-instrument GW events.

  • Disable automated preliminary alerts for all pipelines but gstlal and cWB due to outstanding review items for the other pipelines.

0.4.0 (2019-03-29)

  • This is the penultimate release before LIGO/Virgo observing run 3 (O3).

  • Make detchar results easier to read by formatting as HTML table.

  • Allow iDQ to label DQV onto events based on p(glitch). Adjustable by pipeline.

  • Move functions in tasks/lalinference.py to lalinference_pipe.py in lalsuite.

  • Take into account calibration errors in automatic Parameter Estimation.

  • Do not use margphi option for automatic Parameter Estimation with ROQ waveform since that option is not compatible with ROQ likelihood.

  • Adjust WSGI middleware configuration to adapt to a change in Werkzeug 0.15.0 that broke redirects on form submission in the Flask app. See https://github.com/pallets/werkzeug/pull/1303.

  • Use the new ligo.lw module for reading gstlal’s ranking_data.psd.xml.gz files, because these files are now written using the new LIGO-LW format that uses integer row IDs.

  • Use clean data for parameter estimation.

  • Use production accounting group for PE runs on gracedb events.

  • Change threshold from log-likelihood equals 6 to a dynamic threshold that ensures that all gstlal events uploaded to gracedb get assigned a P_astro value.

0.3.1 (2019-03-18)

  • Fix a bug in translating keys from source_classification.json to keyword arguments for GraceDB.createVOEvent that caused VOEvents to be missing the HasNS and HasRemnant fields.

  • FAR threshold for sending preliminary notices for CBC is changed to 1 per 2 months.

  • Upload log files when LALInference parameter estimation jobs fail or are aborted.

  • Changed the filename source_classification.json to em_bright.json.

  • Change condor log directory from /var/tmp to ~/.cache/condor since gwcelery workers have separate /var/tmp when they are running as condor jobs and that causes problems when gwcelery tries to read log files.

  • Limit the maximum version of gwpy to 0.14.0 in order to work around a unit test failure that started with gwpy 0.14.1. See https://git.ligo.org/emfollow/gwcelery/issues/95.

  • Upload a diff whenever a LIGO/Virgo VOEvent that we receive from GCN does not match the original that we sent.

  • Wait for low-latency or high-latency frame files being transferred to the cluster before parameter estimation starts.

0.3.0 (2019-03-01)

  • Fixed exponent in the expression of foreground count in p_astro_other task.

  • Run the sky map postprocessing and add the PE_READY tag when LALInference finishes.

  • Include EM_COINC triggered circulars to upload to the superevent page.

  • p-astro reads mean values from a file on CIT, new mass-gap category added. Removed redundant functions from p_astro_gstlal module.

  • Continuous deployment on the Caltech cluster now uses a robot keytab and gsissh instead of SSH keys and vanilla ssh because the new my.ligo.org SSH key management does not support scripted access.

  • Improve the isolation between the production and playground instances of GWCelery by deploying them under two separate user accounts on the Caltech cluster.

  • Add functionality for em_bright task to query emfollow/data for trained machine learning classifier and report probabilities based on it.

0.2.6 (2019-02-12)

  • Report an environment tag to Sentry corresponding to the GWCelery configuration module (production, test, playground, or development) in order to differentiate log messages from different deployments.

  • The gwcelery condor command now identifies jobs that it owns by matching both the job batch name and the working directory. This makes it possible to run multiple isolated instances of GWCelery under HTCondor on the same cluster in different working directories.

  • Change the conditions for starting parameter estimation. For every CBC superevent, create an online_pe.ini file suitable for starting LALInference. However, only start LALInference if the false alarm rate is less than once per 2 weeks.

  • Determine PSD segment length for LALInference automatically based on data availability and data quality.

  • Add a Flask-based web interface for manually triggering certain tasks such as sending updated GCN notices.

0.2.5 (2019-02-01)

  • Pass along the GWCelery version number to Sentry.

  • Upload stdout and stderr when dag creation fails and notifications when submitted job fails in Parameter Estimation

  • Allow detchar module’s create_cache to use gwdatafind when frames are no longer in llhoft.

  • The Nagios monitoring plugin will now report on the status of LVAlert subscriptions.

  • Change trials factor to 5 for both CBC and Burst categories. CBC includes the 4 CBC pipelines. Burst includes the 4 searches performed in total by the 2 Burst pipelines. An additional external coincidence search.

  • Automatically set up PE ini file depending on source parameters reported by detection pipelines.

0.2.4 (2018-12-17)

  • Fix broken links in log messages due to changes in GraceDB URL routes.

  • Whenever we send a public VOEvent using GCN, also make the corresponding VOEvent file in GraceDB public.

  • Don’t include Mollweide projection PNG file in VOEvents. The sky map visualizations take longer to generate than the FITS files themselves, so they were unnecessarily slowing down the preliminary alerts.

  • Preliminary GCN FAR threshold is modified to be group (CBC, Burst, Test) specific.

0.2.3 (2018-12-16)

  • Update frame type used in LALInference Parameter Estimation.

  • Handle cases where p_astro_gstlal.compute_p_astro returns NaNs by falling back to p_astro_other.compute_p_astro.

  • Fix a bug that prevented annotations that are specific to 3D sky maps from being performed for multi-resolution FITS files.

  • Fetch the graceid for the new event added from the gracedb logs since superevent packet does not provide information as to which event is added in case of type event_added.

0.2.2 (2018-12-14)

  • Add error handling for nonexistent iDQ frames in detchar module.

0.2.1 (2018-12-14)

  • Update detchar module configuration for ER13.

0.2.0 (2018-12-14)

  • This is the release of GWCelery for ER13.

  • Run two separate instances of Comet, one to act as a broker and one to act as a client. This breaks a cycle that would cause retransmission of GRB notices back to GCN.

  • Fix a race condition that could cause preliminary alerts to be sent out for events for which data quality checks had failed.

  • Unpin the redis package version because recent updates to Kombu and Billiard seem to have fixed the Nagios unit tests.

  • Start the Comet VOEvent broker as a subprocess intead of using multiprocessing and go back to using PyGCN instead of Comet as the VOEvent client. This is a workaround for suspected instability due to a bad interaction between redis-py and multiprocessing.

  • Reset Matplotlib’s style before running ligo-skymap-plot and ligo-skymap-plot-volume. There is some other module (probably in LALSuite) that is messing with the rcparams at module scope, which was causing Mollweide plots to come out with unusual aspect ratios.

  • Run check_vectors upon addition of an event to a superevent if the superevent already has an DQV label.

  • Do not check the DMT-DQ_VECTOR for pipelines which use gated h(t).

  • Remove static example VOEvents from the Open Alert Users Guide. We never used them because activating sample alerts got help until ER13.

  • Disable running the Orchestrator for test events for ER13. After ER13 is over, we need to carefully audit the code and make sure that test events are handled appropriately.

  • Enable public GraceDB entries and public GCNs for mock (MDC) events. For real events in ER13, disable public preliminary GCNs. Instead, advocate signoffs will trigger making events and GCN notices public: ADVOK for initial notices and ADVNO for retraction notices.

  • Include source classification output (BNS/NSBH/BBH/Terrestrial) in GCN Notices.

0.1.7 (2018-11-27)

  • Pin the redis package version at <3 because the latest version of redis breaks the Nagios unit tests.

  • Ditch our own homebrew VOEvent broker and use Comet instead.

  • In addition to traditional flat, fixed-nside sky maps, BAYESTAR will now also upload an experimental multiresolution format described in LIGO-G1800186-v4.

0.1.6 (2018-11-14)

  • Update URL for static example event.

0.1.5 (2018-11-13)

  • Add tasks for submitting HTCondor DAGs.

  • Add a new module, gwcelery.tasks.lalinference, which provides tasks to start parameter estimation with LALInference and upload the results to GraceDB.

  • Depend on lalsuite nightly build from 2018-11-04 to pick up changes to LALInference for Python 3 support.

  • Send static example VOEvents from the Open Alert Users Guide. This will provide a stream of example alerts for astronomers until GraceDB is ready for public access.

  • Add trials factor correction to the event FAR when comparing against FAR threshold to send out preliminary GCN.

  • Require that LIGO/Virgo VOEvents that we receive from GCN match the original VOEvents from GraceDB byte-for-byte, since GCN will now pass through our VOEvents without modification.

0.1.4 (2018-10-29)

  • Work around a bug in astropy.visualization.wcsaxes that affected all-sky plots when Matplotlib’s text.usetex rcparam is set to True (https://github.com/astropy/astropy/issues/8004). This bug has evidently been present since at least astropy 1.3, but was not being triggered until recently: it is likely that some other package that we import (e.g. lalsuite) is now globally setting text.usetex to True.

  • A try except is added around updateSuperevent to handle a bad request error from server side when updating superevent parameters which have nearby values.

  • Send automatic preliminary alerts only for events with a false alarm rate below a maximum value specified by a new configuration variable, preliminary_alert_far_threshold.

  • State vector vetoes will not suppress processing of preliminary sky maps and source classification. They will still suppress sending preliminary alerts.

  • Set open_alert to True for all automated VOEvents.

0.1.3 (2018-10-26)

  • Preliminary GCN is not sent for superevents created from offline gw events.

  • Add dqr_json function to gwcelery.tasks.detchar, which uploads a DQR-compatible json to GraceDB with the results of the detchar checks.

  • Depend on ligo.skymap >= 0.0.17.

  • Fix a bug in sending initial, update, and retraction GCN notices: we were sending the VOEvent filenames instead of the file contents.

0.1.2 (2018-10-11)

  • Setted vetted flag to true for all initial, update, and retraction alerts that are triggered by GraceDB signoffs.

  • Write GraceDB signoffs, instead of just labels, to simulate initial and retraction alerts for mock events, because merely creating the ADVNO or ADVOK label does not cause GraceDB to erase the ADVREQ label. This change makes mock alerts more realistic.

  • Change filename of cWB sky maps from skyprobcc_cWB.fits to cWB.fits.gz for consistency with other pipelines.

  • Any time that we send a VOEvent, first change the GraceDB permissions on the corresponding superevent so that it is visible to the public. Note that this has no effect during the ongoing software engineering runs because LVEM and unauthenticated access are currently disabled in GraceDB.

0.1.1 (2018-10-04)

  • Use the public tag instead of the lvem tag to mark preliminary sky maps for public access rather than LV-EM partner access. Note that GraceDB has not yet actually implemented unauthenticated access, so this should have no effect during our ongoing software engineering runs.

  • Add check_idq function to detchar module, which reads probabilities generated by iDQ.

  • Automated DQV labels should not trigger retraction notices because they prevent preliminary notices from being sent in the first place.

  • The criterion for selecting a superevent’s preferred event now prefers multiple-detector events to single-detector events, with precedence over source type (CBC versus burst). Any remaining tie is broken by using SNR for CBC and FAR for Burst triggers.

  • By default, initial and update alerts will find and send the most recently added public sky map.

  • The initial and update sky maps no longer perform sky map annotations, because they would only be duplicating the annotations performed as part of the preliminary alert.

  • Mock events now include example initial and retraction notices. Two minutes after each mock event is uploaded, there will be either an ADVOK or an ADVNO label applied at random, triggering either an initial or a retraction notice respectively.

  • Depend on ligo-gracedb >= 2.0.1 in order to pull in a bug fix for VOEvents with ProbHasNS or ProbHasRemnant set to 0.0.

  • Use the sentry-sdk package instead of the deprecated raven package for Sentry integration.

0.1.0 (2018-09-26)

  • Separated the external GCN listening handlers into two: one that listens to GCNs about SNEWS triggers and another that listens to Fermi and Swift.

  • Fixed calls to the raven temporal coincidence search so that search results separate SNEWS triggers from Fermi and Swift.

  • Add space-time FAR calculation for GRB and GW superevent coincidences. This only runs when skymaps from both triggers are available to download.

  • Add human vetting for initial GCN notices. For each new superevent that passes state vector checks, the ADVREQ label is applied. Rapid response team users should set their GraceDB notification preferences to alert them on ADVREQ labels. If a user sets the ADVOK label, then an initial notice is issued. If a user sets the ADVNO label, then a retraction notice is issued.

  • Update the LVAlert host for gracedb-playground.ligo.org.

  • Add experimental integration with Sentry for log aggregation and error reporting.

  • Track API and LVAlert schema changes in ligo-gracedb 2.0.0.

0.0.31 (2018-09-04)

  • Refactor external trigger handling to separate it from the orchestrator.

  • Fixed a bug in the VOEvent broker to only issue “iamalive” messages after sending the first VOEvent.

  • Pass group argument to set time windows appropriately when performing raven coincidence searches. Search in the [-600, 60]s range and [-5, 1]s range around external triggers for Burst events and CBC events respectively. Similarly, search in the [-60, 600]s and [-1, 5]s range around Burst and CBC events for external triggers.

  • Compute and upload FAR for GRB external trigger/superevent coincidence upon receipt of the EM_COINC label application to a superevent.

  • Add continuous integration testing for Python 3.7, and run test suite against all supported Python versions (3.6, 3.7).

  • Update ligo.skymap to 0.0.15.

0.0.30 (2018-08-02)

  • Manage superevents for production, test, and MDC events separately.

  • Add some more validation of LIGO/Virgo VOEvents from GCN.

  • Remove now-unused task gwcelery.tasks.orchestartor.continue_if.

  • Add check_vectors run for external triggers.

  • Change the preferred event selection criteria for burst events to be FAR instead of SNR.

  • Add gwcelery nagios subcommand for Nagios monitoring.

  • Incorporate Virgo DQ veto streams into check_vectors

  • Update ligo-raven to 1.3 and ligo-followup-advocate to 0.0.11.

0.0.29 (2018-07-31)

  • Add a workflow graph to superevents module documentation.

  • Add gwcelery condor resubmit as a shortcut for gwcelery condor rm; gwcelery condor submit.

  • Fix deprecation warning due to renaming of ligo.gracedb.rest.Gracedb.createTag to ligo.gracedb.rest.Gracedb.addTag.

  • Update ligo-gracedb to 2.0.0.dev1.

0.0.28 (2018-07-25)

  • Add injection checks to check_vector.

  • Bitmasks are now defined symbolically in detchar.

  • Refactor configuration so that it is possible to customize settings through an environment variable.

0.0.27 (2018-07-22)

  • The preferred event for superevents is now decided based on higher SNR value instead of lower FAR in the case of a tie between groups.

  • A check for the existence of the gstlal trigger database is performed so that compute_p_astro does not return None.

0.0.26 (2018-07-20)

  • Fix spelling of the label that is applied to events after p_astro finishes, changed from P_ASTRO_READY to PASTRO_READY.

  • Run p_astro calculation for mock events.

  • Overhaul preliminary alert pipeline so that it is mostly feature complete for both CBC and Burst events, and uses a common code path for both types. Sky map annotations now occur for both CBC and Burst localizations.

  • Switch to using the pre-registered port 8096 for receiving proprietary LIGO/Virgo alerts on emfollow.ligo.caltech.edu. This means that the capability to receive GCNs requires setting up a site configuration in advance with Scott Barthelmey.

    Once we switch to sending public alerts exclusively, then we can switch back to using port 8099 for anonymous access, requiring no prior site configuration.

0.0.25 (2018-07-19)

  • Reintroduce pipeline-dependent pre/post peeks for check_vector after fixing issue where pipeline information was being looked for in the wrong dictionary.

  • check_vector checks all detectors regardless of instruments used, but only appends labels based on active instruments.

  • Fix a few issues in the GCN broker:

    • Decrease the frequency of keepalive (“iamalive” in VOEvent Transport Protocol parlance) packets from once a second to once a minute at the request of Scott Barthelmey.

    • Fix a possible race condition that might have caused queued VOEvents to be thrown away unsent shortly after a scheduled keepalive packet.

    • Consume and ignore all keepalive and ack packets from the client so that the receive buffer does not overrun.

  • Add p_astro computation for gstlal pipeline. The copmutation is launched for all cbc_gstlal triggers.

0.0.24 (2018-07-18)

  • Revert pipeline-dependent pre/post peeks for check_vector because they introduced a regression: it caused the orchestrator failed without running any annotations.

0.0.23 (2018-07-18)

  • Add timeout and keepalive messages to GCN broker.

  • Update ligo-gracedb to 2.0.0.dev0 and ligo.skymap to 0.0.12.

  • Add superevent duration for gstlal-spiir pipeline.

  • Fix fallback for determining superevent duration for unknown pipelines.

  • Make check_vector pre/post peeks pipeline dependent.

0.0.22 (2018-07-11)

  • Process gstlal-spiir events.

  • Create combined LVC-Fermi skymap in case of coincident triggers and upload to GraceDB superevent page. Also upload the original external trigger sky map to the external trigger GraceDB page.

  • Generalize conditional processing of complex canvases by replacing the continue_if_group_is() task with a more general task that can be used like continue_if(group='CBC').

  • Add a check_vector_prepost configuration variable to control how much padding is added around an event for querying the state vector time series.

    This should have the beneficial side effect of fixing some crashes for burst events, for which the bare duration of the superevent segment was less than one sample.

0.0.21 (2018-07-10)

  • MBTA events in GraceDB leave the search field blank. Work around this in gwcelery.tasks.detchar.check_vectors where we expected the field to be present.

  • Track change in GraceDB JSON response for VOEvent creation.

0.0.20 (2018-07-09)

  • After fixing some minor bugs in code that had not yet been tested live, sending VOEvents to GCN now works.

0.0.19 (2018-07-09)

  • Rewrite the GCN broker so that it does not require a dedicated worker.

  • Send VOEvents for preliminary alerts to GCN.

  • Only perform state vector checks for detectors that were online, according to the preferred event.

  • Exclude mock data challenge events from state vector checks.

0.0.18 (2018-07-06)

  • Add detector state vector checks to the preliminary alert workflow.

0.0.17 (2018-07-05)

  • Undo accidental configuration change in last version.

0.0.16 (2018-07-05)

  • Stop listening for three unnecessary GCN notice types: SWIFT_BAT_ALARM_LONG, SWIFT_BAT_ALARM_SHORT, and SWIFT_BAT_KNOWN_SRC.

  • Switch to SleekXMPP for the LVAlert client, instead of PyXMPP2. Because SleekXMPP has first-class support for publish-subscribe, the LVAlert listener can now automatically subscribe to all LVAlert nodes for which our code has handlers. Most of the client code now lives in a new external package, sleek-lvalert.

0.0.15 (2018-06-29)

  • Change superevent threshold and mock event rate to once per hour.

  • Add gracedb.create_label task.

  • Always upload external triggers to the ‘External’ group.

  • Add rudimentary burst event workflow to orchestrator: it just generates VOEvents and circulars.

  • Create a label in GraceDB whenever em_bright or bayestar completes.

0.0.14 (2018-06-28)

  • Fix typo that was causing a task to fail.

  • Decrease orchestrator timeout to 15 seconds.

0.0.13 (2018-06-28)

  • Change FAR threshold for creation of superevents to 1 per day.

  • Update ligo-followup-advocate to >= 0.0.10. Re-enable automatic generation of GCN circulars.

  • Add “EM bright” classification. This is rudimentary and based only on the point mass estimates from the search pipeline because some of the EM bright classifier’s dependencies are not yet ready for Python 3.

  • Added logic to select CBC events as preferred event over Burst. FAR acts as tie breaker when groups for preferred event and new event match.

  • BAYESTAR now adds GraceDB URLs of events to FITS headers.

0.0.12 (2018-06-28)

  • Prevent receiving duplicate copies of LVAlert messages by unregistering redundant LVAlert message types.

  • Update to ligo-followup-advocate >= 0.0.9 to update GCN Circular text for superevents. Unfortunately, circulars are still disabled due to a regression in ligo-gracedb (see https://git.ligo.org/lscsoft/gracedb-client/issues/7).

  • Upload BAYESTAR sky maps and annotations to superevents.

  • Create (but do not send) preliminary VOEvents for all superevents. No vetting is performed yet.

0.0.11 (2018-06-27)

  • Submit handler tasks to Celery as a single group.

  • Retry GraceDB tasks that raise a TimeoutError exception.

  • The superevent handler now skips LVAlert messages that do not affect the false alarm rate of an event (e.g. simple log messages).

    (Note that the false alarm rate in GraceDB is set by the initial event upload and can be updated by replacing the event; however replacing the event does not produce an LVAlert message at all, so there is no way to intercept it.)

  • Added a query kwarg to superevents method to reduce latency in fetching the superevents from gracedb.

  • Refactored getting event information for update type events so that gracedb is polled only once to get the information needed for superevent manager.

  • Renamed the set_preferred_event task in gracedb.py to update_superevent to be a full wrapper around the updateSuperevent client function. Now it can be used to set preferred event and also update superevent time windows.

  • Many cwb (extra) attributes, which should be floating point numbers, are present in lvalert packet as strings. Casting them to avoid embarassing TypeErrors.

  • Reverted back the typecasting of far, gpstime into float. This is fixed in https://git.ligo.org/lscsoft/gracedb/issues/10

  • CBC t_start and t_end values are changed to 1 sec interval.

  • Added ligo-raven to run on external trigger and superevent creation lvalerts to search for coincidences. In case of coincidence, EM_COINC label is applied to the superevent and external trigger page and the external trigger is added to the list of em_events in superevent object dictionary.

  • cwb and lib nodes added to superevent handler.

  • Events are treated as finite segment window, initial superevent creation with preferred event window. Addition of events to superevents may change the superevent window and also the preferred event.

  • Change default GraceDB server to https://gracedb-playground.ligo.org/ for open public alert challenge.

  • Update to ligo-gracedb >= 1.29dev1.

  • Rename the get_superevent task to get_superevents and add a new get_superevent task that is a trivial wrapper around ligo.gracedb.rest.GraceDb.superevent().

0.0.10 (2018-06-13)

  • Model the time extent of events and superevents using the glue.segments module.

  • Replace GraceDB.get with GraceDB.superevents from the recent dev release of gracedb-client.

  • Fix possible false positive matches between GCNs for unrelated GRBs by matching on both TrigID (which is generally the mission elapsed time) and mission name.

  • Add the configuration variable superevent_far_threshold to limit the maximum false alarm rate of events that are included in superevents.

  • LVAlert handlers are now passed the actual alert data structure rather than the JSON text, so handlers are no longer responsible for calling json.loads. It is a little bit more convenient and possibly also faster for Celery to deserialize the alert messages.

  • Introduce Production, Development, Test, and Playground application configuration objects in order to facilitate quickly switching between GraceDB servers.

  • Pipeline specific start and end times for superevent segments. These values are controlled via configuration variables.

0.0.9 (2018-06-06)

  • Add missing LVAlert message types to superevent handler.

0.0.8 (2018-06-06)

  • Add some logging to the GCN and LVAlert dispatch code in order to diagnose missed messages.

0.0.7 (2018-05-31)

  • Ingest Swift, Fermi, and SNEWS GCN notices and save them in GraceDB.

  • Depend on the pre-release version of the GraceDB client, ligo-gracedb 1.29.dev0, because this is the only version that supports superevents at the moment.

0.0.6 (2018-05-26)

  • Generate GCN Circular drafts using ligo-followup-advocate.

  • In the continuous integration pipeline, validate PEP8 naming conventions using pep8-naming.

  • Add instructions for measuring test coverage and running the linter locally to the contributing guide.

  • Rename gwcelery.tasks.voevent to gwcelery.tasks.gcn to make it clear that this submodule contains functionality related to GCN notices, rather than VOEvents in general.

  • Rename gwcelery.tasks.dispatch to gwcelery.tasks.orchestrator to make it clear that this module encapsulates the behavior associated with the “orchestrator” in the O3 low-latency design document.

  • Mock up calls to BAYESTAR in test suite to speed it up.

  • Unify dispatch of LVAlert and GCN messages using decorators. GCN notice handlers are declared like this:

    import lxml.etree
    from gwcelery.tasks import gcn
    
    @gcn.handler(gcn.NoticeType.FERMI_GBM_GND_POS,
                 gcn.NoticeType.FERMI_GBM_FIN_POS)
    def handle_fermi(payload):
        root = lxml.etree.fromstring(payload)
        # do work here...
    

    LVAlert message handlers are declared like this:

    import json
    from gwcelery.tasks import lvalert
    
    @lvalert.handler('cbc_gstlal',
                     'cbc_pycbc',
                     'cbc_mbta')
    def handle_cbc(alert_content):
        alert = json.loads(alert_content)
        # do work here...
    
  • Instead of carrying around the GraceDB service URL in tasks, store the GraceDB host name in the Celery application config.

  • Create superevents by simple clustering in time. Currently this is only supported by the gracedb-dev1 host.

0.0.5 (2018-05-08)

  • Disable socket access during most unit tests. This adds some extra assurance that we don’t accidentally interact with production servers during the unit tests.

  • Ignore BAYESTAR jobs that raise a DetectorDisabled error. These exceptions are used for control flow and do not constitute a real error. Ignoring these jobs avoids polluting logs and the Flower monitor.

0.0.4 (2018-04-28)

  • FITS history and comment entries are now displayed in a monospaced font.

  • Adjust error reporting for some tasks.

  • Depend on newer version of ligo.skymap.

  • Add unit tests for the gwcelery condor submit subcommand.

0.0.3 (2018-04-27)

  • Fix some compatibility issues between the gwcelery condor submit subcommand and the format of condor_q -totals -xml with older versions of HTCondor.

0.0.2 (2018-04-27)

  • Add gwcelery condor submit and related subcommands as shortcuts for managing GWCelery running under HTCondor.

0.0.1 (2018-04-27)

  • This is the initial release. It provides rapid sky localization with BAYESTAR, sky map annotation, and sending mock alerts.

  • By default, GWCelery is configured to listen to the test LVAlert server.

  • Sending VOEvents to GCN/TAN is disabled for now.