Release notes v3.14

The release 3.14 is the result of 3 months of work involving nearly 200 pull requests. The major highlight is the complete vectorization of hazardlib.

The complete list of changes is listed in the changelog:

https://github.com/gem/oq-engine/blob/engine-3.14/debian/changelog

Work on hazardlib

One year ago we started a vectorization program for the GMPEs in hazardlib, described here:

https://github.com/gem/oq-engine/blob/engine-3.14/doc/breaking-hazardlib.md

As of engine 3.14 all of the ~700 GMPEs in hazardlib are vectorized, i.e. they use the new API in terms of numpy arrays and not Python objects. The new and better way to work with GMPEs is documented here:

https://docs.openquake.org/oq-engine/advanced/developing.html#working-with-gmpes-directly-the-contextmaker

As a consequence of the vectorization, classical calculations can be much faster than before (of the order of 2-10 times faster) in the case of few sites. Models with GMPEs based on HDF5 tables - like the Canada and USA models - have been specially optimized, so that you can have significant speedups - in the order of 3-5 times for the Canada 2015 model - even in the many-sites case.

There was also a substantial amount of work for collapsing the context objects, even for nonparametric sources: that will give significant speedups in future versions of the engine.

Also, numba is used in a few more hotspots, giving significant speedups in some classical calculations.

There were some improvements to KiteSurfaces and to MultiFaultSources. In particular the performance of MultiFaultSources in the engine was disastrous: all the time was spent in computing the bounding box and the encircling polygon. This is now fixed.

The magnitudes in multiFaultSources have been rounded to two digits after the decimal point.

We added a check on geo.Line forbidden degenerate lines with less than two points.

We extended the ModifiableGMPE to work with GMPETable subclasses, with the NRCan15 site term and with the Al Atik Sigma model.

We introduced the concept of computed rupture parameters, which was used to vectorize the McVerry GMPEs.

Prajacta Jadhav and Dharma Wijewickreme contributed the GMPE Youd et al. (2002) and the corresponding site parameters.

Tom Son contributed a few bug fixes to the GMPE Kuehn et al. (2020).

Finally, we raise an error when the total standard deviation is zero and the truncation level is not zero (currently this may happen only with the GMPE YenierAtkinson2015BSSA).

New hazard features

The UCERF calculator has been removed. Now you can run an UCERF calculation by using the regular classical calculator, provided you have the UCERF model in a new format using multiFaultSources. The new format has substantial advantages, for instance it is possible to combine the UCERF model with other models.

We implemented a cache for the distances that speedups substantially the calculation (we are talking about a 2x speedup). Since the distance calculation algorithm has changed internally, the numbers are expected to be slightly different than in past versions.

At user request, it is now possible to disaggregate by all realizations by setting in the job.ini file

num_rlzs_disagg = 0

Naturally, if you have a large number of realizations that can cause your calculation to run out of memory or to become extra-slow. Use this feature with care.

New risk features

At user request, we implemented multi-tag aggregation: it is now possible to aggregate across different tag combinations within a single calculation. The exporters have been changed accordingly to export a file for each tag combination. For instance setting in the job.ini something like

aggregate_by = NAME_1, taxonomy; NAME_1; taxonomy

would perform four different kinds of aggregations (by “NAME_1 & taxonomy”, by “NAME_1”, by “taxonomy” and total aggregation) and the aggregate risk and aggregate curves exporters would export four files each. The export format is the same, however while before the totals were exported in the same file, now they are exported in a separate file.

We also changed how the Aggregate Risk output is computed, reverting back to the definition of previous versions of the engine, directly related to the average annual losses.

Bug fixes

We discovered a bug in the check on unique section IDs for multiFaultSources, such that non-unique sections IDs were possible. Also, section IDs were incorrectly recognized as source IDs. Both issues are now fixed.

job.ini files containing a Byte Order Mark (BOM) where not read correctly, causing issues particularly on the Windows platform.

There was a small bug with discard_trts, such that in some cases too many TRTs were discarded. This is now fixed.

Some event_based_damage calculations could fail in a cluster environment due to the parent_dir parameter being passed incorrectly.

The web API /v1/calc/run was failing when uploading files due to a sorting issue.

We changed the web API /extract/events to return events sorted by ID: this avoids an error with the QGIS plugin when visualizing GMFs.

There was an error in the aggrisk exporter not considering the realizations generating zero losses and therefore computing incorrectly the means.

There was a bug when exporting Asset Loss Maps in some situations, due to a 32 bit/64 bit mismatch in the conditional_loss_ratio parameter.

We improved the error checking for calculations with site amplification.

We fixed the --hc feature that was failing when starting from a preclassical calculation.

oq commands

We removed the obsolete commands oq to_shapefile and oq from_shapefile and the explicit dependency from pyshp.

We changed oq zip <directory> to generate a single big archive instead of an archive for each job.ini.

We extended oq postzip to post multiple files at once.

We extended oq reduce_sm to work with multiFaultSources.

We made oq engine --multi --run more robust against out-of-memory errors.

IT changes

The support for Python 3.6, deprecated in version 3.13, has been removed. The engine should still work with it but we do not test Python 3.6 anymore and can stop working at any moment without warnings. On the other hand, Python 3.9 is now officially supported.

We upgraded Shapely to version 1.8.0. This is a big change, since the new version is incompatible with the past and produces different numbers: as a consequence the engine produces slightly different results than in previous versions.

We updated Django to release 3.2.12 to avoid security bugs.

We changed some internals to make it easier to run the engine in cluster environments with a shared database.

We made it possible to set a default pointsource_distance in the global configuration file openquake.cfg.

Finally the install.py script has been improved:

• the server installation correctly installs the systemctl files both on Debian-based and RedHat-based distributions.

• we added a missing [directory] section in the openquake.cfg file generated by install.py.

• we now raise a helpful error message when ensurepip is missing.

Other

The classical calculator now stores unique rupture IDs in the rup DataFrame. There are some limitations, and in particular you cannot expect the same calculation on different machine to produce the same IDs unless you fix the parameter concurrent_tasks too.

We removed the applyToSourceType functionality of the logic tree, since it was not used and broken anyway.

We raised the limit to 94 GMPEs per tectonic region type and 94 branches per source model logic tree.

In the ruptures dataset for event based calculations we now store the occurrence rate multiplied by the source model weight, to simplify some consistency checks.

Finally, as usual, we worked at improving the documentation.