Edgar (a.k.a the AP03 project) exists to visualize bird occurrence records and distribution maps. Occurrence records are basically coordinates where a specific species of bird was seen, and these coordinates are used to generate the distribution maps. Therefor, the quantity and quality of the occurrence records is very important to the usefulness of the project. To obtain a large number of quality occurrences, Edgar will be using the data from the Atlas of Living Australia (ALA). It was decided that we will keep a local copy of all bird occurrence records from ALA, because using all of the data live (i.e. straight from the ALA servers) isn’t feasible for a few reasons:

  • Edgar needs to handle occurrence records from multiple sources.
  • Using ALA directly would make Edgar more difficult for other people to set up and use with different occurrence record sources.
  • The rating/vetting categories in Edgar (assumed valid, known valid, assumed invalid, and known invalid) do not correspond very well with the ALA system of assertions.
  • Server-side clustering of records is required for performance reasons. ALA provides clustering in the form of map layers, but does not provide the ability to colour the layers depending on our specific vetting categories. If Edgar performed the clustering itself using live records from ALA servers, it would not run fast enough to display to the user (could take in excess of five minutes for a single map).

Keeping a local copy of the occurrence records poses its own problems, however.

Disk Size

Initial testing indicates that we can keep each record under 100 bytes in a MySQL database. ALA has around 18 million occurrence records for birds, so this means that storing all of the occurrences will require less than two gigabytes of storage on disk. The size itself shouldn’t be a problem in this case, but the performance of database that size remains untested.

Speed

Importing 18 million records at once will take a considerable amount of time. Instead of repeatedly downloading all records when Edgar needs an update, we plan on doing incremental updates. This means that there will be an initial import of all 18 million records, but subsequent imports will only download records that have been added, deleted, or modified. This will allow Edgar to synchronise its data with ALA more frequently.

Initial testing indicates that we can download records from ALA at roughly 500 to 2000 records per second using a concurrent architecture for HTTP requests. This means the initial import of 18 million records should take between 3 and 10 hours.

Species Changes

The list of bird species will change as new species are found, and existing species are merged together or split apart. As part of the ALA syncing process, these additions, splits and merges will be incorporated into the local database.

Connectivity and Availability Problems

Because an import can run for hours, it’s highly likely that ALA web services may become unavailable while an import is running. This could happen for many reasons, including ALA servers being restarted or temporary network connectivity problems. The importer must accomodate these disruptions without producing incorrect results.

When a request to ALA fails for any reason, the request is retried using an exponential back off algorithm. With the current settings, if a request fails it will wait 10 seconds before retrying the request. If the retry fails, then the wait duration is doubled before performing the next retry. The fifth and last retry will happen over five minutes after the first try, and if that fails then the import will stop running and log an error.

Five retries over five minutes should be sufficient to overcome common availability problems, but if there are recurring import failures it may require more sophisticated error handling.

Data Quality

ALA aggregates occurrence records from many different sources, so the data quality varies. Some occurrences are too old to be reliable, such as the 158 records for the extinct Tasmanian Tiger. Some occurrences contain inaccurate coordinates, or no coordinates at all. Some occurrences have been obfuscated (made inaccurate on purpose) to hide the location of endangered species. ALA also keeps records of preserved specimens, images, sounds, and genetic information, which are not relevant to Edgar.

ALA has a comprehensive system of assertions that indicate known problems with each occurrence record. For instance, there are assertions for the latitude and longitude being swapped, and for coordinates of terrestrial animals that appear in the ocean. Most of the assertions come from the data cleaning efforts of ALA, but some of them also come from users reporting issues with individual records.

The import will ignore irrelevant records, and try to assign assumed valid or assumed invalid ratings to each occurrence based on the ALA assertions that are present. The exact combination of assertions that will cause a record to be marked as “assumed invalid” has not been decided upon yet, but we are currently using reasonable defaults.

We hope to get access to the accurate coordinates for obfuscated occurrences. The obfuscated occurrences may or may not be shown on the map, but the accurate coordinates will not be display. The accurate coordinates will only be used to generate distribution maps, which will not contain any individual occurrence coordinates when displayed.

Implementation

The syncing will be performed by python scripts. These are currently available from https://github.com/jcu-eresearch/Edgar/tree/master/importing

A cron job will be set up to run a single script, which is currently called ala_db_update.py. There is a configuration file that contains database connection settings and options to control minor behaviours of the importer, such as how much information is logged while it runs.

The syncing process roughly involves these steps:

  1. Record the start time for the import
  2. Get the last import time from the database
  3. Query ALA for a list of all bird species
  4. Update the species in the database
  5. Query ALA for occurrence records that have been changed or added since the last import time
  6. Add and update occurrences in the database
  7. Delete occurrences from the database that no longer exist at ALA
  8. Check that the number of occurrences in the local database matches the number of occurrences at ALA
  9. Update the last import time in the database to the time that the import started

If the import fails, the last import time will not be updated in the database. This will ensure that the next run of the importer will get all of the occurrences that the failed run did not.



blog comments powered by Disqus