Release notes v3.0

This release drops support for Python 2.7. From now on, the engine requires Python 3.5+ to run. Apart from this change, several new features were implemented, like the support for time dependent source models, automatic gridding of the exposure, enhancements to the disaggregation calculator, improvements to the task distribution and a lot more. Over 110 issues were closed. For the complete list of changes, please see the changelog: https://github.com/gem/oq-engine/blob/engine-3.0/debian/changelog .

Management of the source model

As of engine 3.0, the sources are split upfront, before filtering. This approach became possible after optimizing the splitting procedure for fault sources which now is extremely fast. It has several benefits in terms of simplicity of the code and better estimation of the task computational weight.

Moreover, it is now possible to warn the user upfront, if is using discretization parameters such as the area_source_discretization, the rupture_mesh_spacing and the complex_fault_mesh_spacing which are too small, thus producing millions of ruptures without need (this is a common mistake).

Complex fault sources are now split in more sub-sources than before and this produces a substantial improvement of the task distribution. It also fixed a bug with the event_based_rupture calculator generating too few tasks.

A different bug with the classical calculator generating too few tasks when the option optimize_same_id_sources is set has been fixed as well.

We extended the check on duplicated IDs in the source model to models in the NRML 0.5 format. This means that if a single source model is split in several files (i.e. the <uncertaintyModel> tag in the source model logic tree file contains several paths) the source IDs must be unique across all files.

Source models with time-dependent sources now require two new tags: investigation_time (mandatory) and start_time (optional, but it will likely become mandatory in the future). The investigation_time tag is used to check the investigation_time parameter in the job.ini file, so that the user cannot accidentally use the wrong investigation time.

Now we log some information about the floating/spinning factors, which are relevant for point sources and area sources (see the manual section 2.1.1.1 for an explanation). This is useful for us since in the future we may introduce some optimization to reduce the floating/spinning of the ruptures. Users can simply ignore such logs.

Hazard

We extended the event based calculator to work with mutually exclusive (mutex) sources: this is relevant for the Japan model and others. Thanks to some fixes to the GriddedRupture class in hazardlib now it is also possible to export ruptures of this kind, which are relevant for several recent hazard models.

We fixed how the event year is set in event based calculations. Before the events sets were considered temporally ordered, i.e. in a case with 2 stochastic event sets (SES) with an investigation time of 50 years one could have years (0, 4, 13, 27, 49) in the first SES and (55, 61, 90) in the second SES: now we would have (0, 4, 13, 27, 49) in the first SES and (5, 11, 40) in the second. The event in the second SES have no greater years that the events in the first SES since each event set starts from the year 0. This is the correct way of proceeding in the case of time-dependent models, which were not supported before.

The net effect of the change is that the ordering of the event loss table can be different than before, since the year was used (solely) as an ordering parameter in the exporter.

Hazard disaggregation

We implemented statistical disaggregation outputs. This is implemented in a straightforward way: if there are multiple realizations and if in the job.ini file the parameters mean_hazard_curves and/or quantile_hazard_curves are set, then the mean and/or quantiles of the hazard outputs are computed. You can export such outputs as usual.

The parameter disagg_outputs is now honored: for instance if you have in the job.ini disagg_outputs = Mag Mag_Dist, then only the outputs of kind Mag and Mag_Dist are computed and stored. Before all of them were computed and stored and the parameter affected only the export. If disagg_outputs is not given, all of the available disaggregation outputs are generated, as in the past.

We introduced, experimentally, a disaggregation by source feature. It is restricted to the case of a single realization. In that case, if you set disagg_by_src=true in the job.ini, then an output “Disaggregation by Source” is generated. When exported, you get a set of .csv files with fields (source_id, source_name, poe). For each source with nonzero contribution the contribution to the total probability of exceedence is given.

Finally, we fixed a small bug in the disaggregation calculator with parameter poes_disagg: we were reading the source model twice without reason.

Hazardlib/HMTK/SMTK

We optimized the Yu et al. (2013) GMPEs for China which is now several time faster than before.

Graeme Weatherill ported to Python 3.5 the Strong Motion Toolkit, which depends on hazardlib and is a part of the OpenQuake suite.

Nick Ackerley fixed a bug in the HMTK plotting libraries and added the ability to customize the figure size.

The source writer in hazardlib now checks that the sum of the probabilities of occurrence is 1, when saving nonparametric sources. This avoids errors when building time-dependent models.

Risk

The management of the exposure has been refactored and improved. It is now possible to run a risk calculation from a pre-imported exposure. This is important because the engine is powerful enough to run calculations with millions of assets and it is convenient to avoid reimporting the exposure every time if it does not change.

On the same note, it is possible to use a pre-imported risk model, without having to reimport it at each risk calculation.

As of engine 3.0, the exposure should contain a tag <occupancyPeriod> listing the occupancy periods, i.e. subsets of day, night, transit, including the case of no occupancy periods. If such tag is missing, you will get a warning. If such tag is present, but the listed occupancy periods are inconsistent with the ones found in the assets, a clear error is raised.

The ability to import CSV exposures has been extended to the cases when there are occupancy periods, which are managed simply as additional fields. Insurance parameters (insured_losses/deductibles) are still not supported in CSV format and the XML format is still needed for that case. We plan to keep working on that in the future.

We extended the engine logic to read the sites from the hazard curves, if available. Moreover we changed the logic to extract the sites from the exposure in precedence over the site model, if the sites are not provided explicitly (via a sites.csv file, the hazard curves or a region).

This was necessary because of a new feature, i.e. the automatic gridding of the exposure. If your job.ini file contains an exposure, no region parameter and the parameter region_grid_spacing, then a grid of sites is automatically generated from the region encircling the exposure (if there is a region parameter the grid is generated from the region as before).

Automatic gridding of the exposure is very important, because often the assets are provided with a very high resolution (say 1 km); by providing a coarser grid (say 5 km) the hazard part of the calculation can become a lot faster (say 25 times faster) while producing very similar results for the aggregated losses.

Care is taken so that points of the grid with no close assets (i.e. outside the grid spacing) are automatically discarded; moreover, there are checks to make sure that all assets are associated to the grid.

Event Based Risk calculations with sampling are now officially supported, after years in which this feature was experimental and not really working. This is relevant for cases like the India model were the number of realizations is so large (there are over 200,000 realizations) that full enumeration is not an option and sampling of the logic tree is a necessity.

WebAPI/WebUI/QGIS plugin

We fixed some permission bugs with the WebUI when groups are involved: now it is possible to download the outputs of calculations run by other people in the same group. This is useful for organizations wanting to enable the authentications features of the WebUI. By default, as always, the WebUI is public.

We added more risk outputs to the extract API. In particular now it is possible to extract also the losses by asset coming from event based risk and classical risk calculations. Moreover it is possible to aggregate such losses by using the usual aggregation interfaces (the web API and the QGIS plugin).

Bug fixes/additional checks

Running a calculation on a big machine, copying the datastore on a small machine (i.e. a laptop) and exporting the results now works even for calculations involving GMPE tables i.e. prediction equations implemented as numeric tables in .hdf5 format. This is relevant for the Canada model and others.

We have now a better error message when there are duplicated sites in the CSV; in particular, the first line with the duplicates is shown, as well as the line number.

We fixed a bug with newlines in the logic tree path breaking the CSV exporter for the realizations output. This was happening with models split in multiple files, like the CCARA model.

We fixed a bug in classical_risk, happening when the number of statistics was larger than the number of realizations (for instance, in a case with two realizations, computing mean, quantile-0.15 and quantile-0.85, i.e. three statistics).

We fixed the strange issue of very small negative numbers appearing in scenario_damage outputs: this happened due to rounding errors. Now the correct result (i.e. zeros) is stored.

We added a check in calculations reading the GMFs in CSV format: now there must be a single realization in the input file.

When running a scenario calculation using precomputed GMFs, a clear error message appears when the IMTs in the GMFs are not compatible with the IMTs in the fragility/vulnerability file.

We added a check against duplicated fields in the exposure CSV.

oq commands

The command oq info has been extended to source model logic tree files: in that case it reports a summary for the full composite source model.

The command oq dbserver stop and oq workers stop now correctly stops the zmq workers (relevant for the experimental zmq mode).

The command oq show was not working in multiuser situations for calculations with a parent, since the parent was being read from the wrong directory. This issue has been fixed.

The command oq show job_info now returns the amount of data received from the controller node while the computation is running: before this information was available only at the end of the computation.

There is a new command oq importcalc host calc_id username password to import remote calculations into the local engine database. The command has some limitations: it works only for calculations without a parent and only if there are no clashes between the remote calculation ID and the local calculation ID. It should be considered an internal command.

Internals

A huge improvement has been made in cluster situations: now the results are returned via ZeroMQ and not via rabbimq. This allows us to bypass the limitations of rabbitmq: large computations can be run without running out of disk space in the mnesia directory. Hundreds of thousands of tasks can be generated without issues, a feat previously impossible.

Notice that you may need to adjust the master node firewall, if already configured, to allow incoming traffic on TCP port range 1907-1920.

Now we use the port 1907 for the DbServer, when installing the engine from the packages. When installing from sources, the port is the number 1908, as before. In this way an installation from packages can coexists with an installation from sources out of the box.

The task distribution code has been simplified and features in the experimental/proof-of-concept state has been removed: in particular the support to ipython and the support to SGE have disappeared. They were not used and they were a significant maintenance cost. The default for the distribution parameter in the configuration file openquake.cfg is now processpool, not futures. The old syntax is still supported, but a warning will be printed, saying to use processpool instead. Technically we do not rely anymore on the Python module concurrent.futures, we use the module multiprocessing directly.

The engine now stores more information about the sources. In particular in the case of event based calculations the source_info dataset contains the number of events generated by each source. Moreover, there is an utility utils/reduce_sm than can read such information and reduce a source model by removing all sources not producing events.

As usual a lot of refactoring was done and several engine tests are faster than before. Also the event based risk demo is several times faster than before.

Deprecations/removals

The engine does not work anymore with Python 2. Hazardlib and the Hazard Modeller Toolkit, included in the engine, still work with Python 2 but this is only incidental: they may stop working at any moment without warning, since we are not testing anymore the engine libraries with Python 2.

The old commands oq engine --run-hazard and oq engine --run-risk, deprecated two years ago, have been finally removed. The only command to use to run calculations is oq engine --run, without distinction between hazard and risk.

The function openquake.hazardlib.calc.stochastic.stochastic_event_set has been deprecated: you can use the function openquake.hazardlib.calc.stochastic.sample_ruptures instead.

As usual, exporting the results of a calculation executed with a previous version of the engine is not supported, except for hazard curves/maps/spectra. We recommend first to export all of the results you need and then to upgrade to the latest version of the engine.