XLoader - ckanext-xloader

Loads CSV (and similar) data into CKAN's DataStore. Designed as a replacement for DataPusher because it offers ten times the speed and more robustness (hence the name, derived from "Express Loader")

OpenGov Inc. has sponsored this development, with the aim of benefitting open data infrastructure worldwide.

Key differences from DataPusher

Speed of loading

DataPusher - parses CSV rows, converts to detected column types, converts the data to a JSON string, calls datastore_create for each batch of rows, which reformats the data into an INSERT statement string, which is passed to PostgreSQL.

XLoader - pipes the CSV file directly into PostgreSQL using COPY.

In tests, XLoader is over ten times faster than DataPusher.

Robustness

DataPusher - one cause of failure was when casting cells to a guessed type. The type of a column was decided by looking at the values of only the first few rows. So if a column is mainly numeric or dates, but a string (like "N/A") comes later on, then this will cause the load to error at that point, leaving it half-loaded into DataStore.

XLoader - loads all the cells as text, before allowing the admin to convert columns to the types they want (using the Data Dictionary feature). In future it could do automatic detection and conversion.

Simpler queueing tech

DataPusher - job queue is done by ckan-service-provider which is bespoke, complicated and stores jobs in its own database (sqlite by default).

XLoader - job queue is done by RQ, which is simpler, is backed by Redis, allows access to the CKAN model and is CKAN's default queue technology (sinc CKAN 2.7). You can also debug jobs easily using pdb. Job results are stored in Sqlite by default, and for production simply specify CKAN's database in the config and it's held there - easy.

(The other obvious candidate is Celery, but we don't need its heavyweight architecture and its jobs are not debuggable with pdb.)

Separate web server

DataPusher - has the complication that the queue jobs are done by a separate (Flask) web app, apart from CKAN. This was the design because the job requires intensive processing to convert every line of the data into JSON. However it means more complicated code as info needs to be passed between the services in http requests, more for the user to set-up and manage - another app config, another apache config, separate log files.

XLoader - the job runs in a worker process, in the same app as CKAN, so can access the CKAN config, db and logging directly and avoids many HTTP calls. This simplification makes sense because the xloader job doesn't need to do much processing - mainly it is streaming the CSV file from disk into PostgreSQL.

Caveat - column types

Note: With XLoader, all columns are stored in DataStore's database as 'text' type (whereas DataPusher did some rudimentary type guessing - see 'Robustness' above). However once a resource is xloaded, an admin can use the resource's Data Dictionary tab (CKAN 2.7 onwards) to change these types to numeric or datestamp and re-load the file. When migrating from DataPusher to XLoader you can preserve the types of existing resources by using the migrate_types command.

There is scope to add functionality for automatically guessing column type - offers to contribute this are welcomed.

Requirements

Works with CKAN 2.7.x and later.

Works with CKAN 2.3.x - 2.6.x if you install ckanext-rq.

Compatibility with core CKAN versions:

CKAN version	Compatibility
2.3	yes, but no longer tested and you must install ckanext-rq
2.4	yes, but no longer tested and you must install ckanext-rq
2.5	yes, but no longer tested and you must install ckanext-rq
2.6	yes, but no longer tested and you must install ckanext-rq
2.7	yes
2.8	yes
2.9	not yet - see ckan/ckanext-xloader#104

Installation

To install XLoader:

1. Activate your CKAN virtual environment, for example:

   . /usr/lib/ckan/default/bin/activate
   

2. Install the ckanext-xloader Python package into your virtual environment:

   pip install ckanext-xloader
   

3. Install dependencies:

   pip install -r https://raw.githubusercontent.com/ckan/ckanext-xloader/master/requirements.txt
   pip install -U requests[security]
   

4. If you are using CKAN version before 2.8.x you need to define the populate_full_text_trigger in your database

   sudo -u postgres psql datastore_default -f full_text_function.sql
   

   If successful it will print

   CREATE FUNCTION
   ALTER FUNCTION
   

   NB this assumes you used the defaults for the database name and username. If in doubt, check your config's ckan.datastore.write_url. If you don't have database name datastore_default and username ckan_default then adjust the psql option and full_text_function.sql before running this.

5. Add xloader to the ckan.plugins setting in your CKAN config file (by default the config file is located at /etc/ckan/default/production.ini).

   You should also remove datapusher if it is in the list, to avoid them both trying to load resources into the DataStore.

   Ensure datastore is also listed, to enable CKAN DataStore.

6. If it is a production server, you'll want to store jobs info in a more robust database than the default sqlite file. It can happily use the main CKAN postgres db by adding this line to the config, but with the same value as you have for sqlalchemy.url:

   ckanext.xloader.jobs_db.uri = postgresql://ckan_default:pass@localhost/ckan_default
   

   (This step can be skipped when just developing or testing.)

7. Restart CKAN. For example if you've deployed CKAN with Apache on Ubuntu:

   sudo service apache2 reload
   

8. Run the worker. First test it on the command-line:

   paster --plugin=ckan jobs -c /etc/ckan/default/ckan.ini worker
   

   or if you have CKAN version 2.6.x or less (and are therefore using ckanext-rq):

   paster --plugin=ckanext-rq jobs -c /etc/ckan/default/ckan.ini worker
   

   Test it will load a CSV ok by submitting a CSV in the web interface or in another shell:

   paster --plugin=ckanext-xloader xloader submit <dataset-name> -c /etc/ckan/default/ckan.ini
   

   Clearly, running the worker on the command-line is only for testing - for production services see:

   http://docs.ckan.org/en/ckan-2.7.0/maintaining/background-tasks.html#using-supervisor

   If you have CKAN version 2.6.x or less then you'll need to download supervisor-ckan-worker.conf and adjust the command to reference ckanext-rq.

Config settings

Configuration:

# The connection string for the jobs database used by XLoader. The
# default of an sqlite file is fine for development. For production use a
# Postgresql database.
ckanext.xloader.jobs_db.uri = sqlite:////tmp/xloader_jobs.db

# The formats that are accepted. If the value of the resource.format is
# anything else then it won't be 'xloadered' to DataStore (and will therefore
# only be available to users in the form of the original download/link).
# Case insensitive.
# (optional, defaults are listed in plugin.py - DEFAULT_FORMATS).
ckanext.xloader.formats = csv application/csv xls application/vnd.ms-excel

# The maximum size of files to load into DataStore. In bytes. Default is 1 GB.
ckanext.xloader.max_content_length = 1000000000

# To always use messytables to load data, instead of attempting a direct
# PostgreSQL COPY, set this to True. This more closely matches the
# DataPusher's behavior. It has the advantage that the column types
# are guessed. However it is more error prone, far slower and you can't run
# the CPU-intensive queue on a separate machine.
ckanext.xloader.just_load_with_messytables = False

# The maximum time for the loading of a resource before it is aborted.
# Give an amount in seconds. Default is 60 minutes
ckanext.xloader.job_timeout = 3600

# Ignore the file hash when submitting to the DataStore, if set to True
# resources are always submitted (if their format matches), if set to
# False (default), resources are only submitted if their hash has changed.
ckanext.xloader.ignore_hash = False

# When loading a file that is bigger than `max_content_length`, xloader can
# still try and load some of the file, which is useful to display a
# preview. Set this option to the desired number of lines/rows that it
# loads in this case.
# If the file-type is supported (CSV, TSV) an excerpt with the number of
# `max_excerpt_lines` lines will be submitted while the `max_content_length`
# is not exceeded.
# If set to 0 (default) files that exceed the `max_content_length` will
# not be loaded into the datastore.
ckanext.xloader.max_excerpt_lines = 100

Developer installation

To install XLoader for development, activate your CKAN virtualenv and in the directory up from your local ckan repo:

git clone https://github.com/ckan/ckanext-xloader.git
cd ckanext-xloader
python setup.py develop
pip install -r requirements.txt
pip install -r dev-requirements.txt

Upgrading from DataPusher

To upgrade from DataPusher to XLoader:

1. Install XLoader as above, including running the xloader worker.

2. (Optional) For existing datasets that have been datapushed to datastore, freeze the column types (in the data dictionaries), so that XLoader doesn't change them back to string on next xload:

   paster --plugin=ckanext-xloader migrate_types -c /etc/ckan/default/ckan.ini
   

3. If you've not already, change the enabled plugin in your config - on the ckan.plugins line replace datapusher with xloader.

4. (Optional) If you wish, you can disable the direct loading and continue to just use messytables - for more about this see the docs on config option: ckanext.xloader.just_load_with_messytables

5. Stop the datapusher worker:

   sudo a2dissite datapusher
   

6. Restart CKAN:

   sudo service apache2 reload
   sudo service nginx reload
   

Command-line interface

You can submit single or multiple resources to be xloaded using the command-line interface.

e.g.

paster --plugin=ckanext-xloader xloader submit <dataset-name> -c /etc/ckan/default/ckan.ini

For debugging you can try xloading it synchronously (which does the load directly, rather than asking the worker to do it) with the -s option:

paster --plugin=ckanext-xloader xloader submit <dataset-name> -s -c /etc/ckan/default/ckan.ini

See the status of jobs:

paster --plugin=ckanext-xloader xloader status -c /etc/ckan/default/development.ini

Submit all datasets' resources to the DataStore:

paster --plugin=ckanext-xloader xloader submit all -c /etc/ckan/default/ckan.ini

Re-submit all the resources already in the DataStore (Ignores any resources that have not been stored in DataStore e.g. because they are not tabular):

paster --plugin=ckanext-xloader xloader submit all-existing -c /etc/ckan/default/ckan.ini

Full list of XLoader CLI commands:

paster --plugin=ckanext-xloader xloader --help

Jobs and workers

Main docs for managing jobs: <https://docs.ckan.org/en/latest/maintaining/background-tasks.html#managing-background-jobs>

Main docs for running and managing workers are here: https://docs.ckan.org/en/latest/maintaining/background-tasks.html#running-background-jobs

Useful commands:

Clear (delete) all outstanding jobs:

paster --plugin=ckan jobs clear [QUEUES] -c /etc/ckan/default/development.ini

If having trouble with the worker process, restarting it can help:

sudo supervisorctl restart ckan-worker:*

Troubleshooting

KeyError: "Action 'datastore_search' not found"

You need to enable the datastore plugin in your CKAN config. See 'Installation' section above to do this and restart the worker.

ProgrammingError: (ProgrammingError) relation "_table_metadata" does not exist

Your DataStore permissions have not been set-up - see: <https://docs.ckan.org/en/latest/maintaining/datastore.html#set-permissions>

When editing a package, all its existing resources get re-loaded by xloader

This behavior was documented in Issue 75 and is related to a bug in CKAN that is fixed in versions 2.6.9, 2.7.7, 2.8.4 and 2.9.0+.

Running the Tests

The first time, your test datastore database needs the trigger applied:

sudo -u postgres psql datastore_test -f full_text_function.sql

To run the tests, do:

nosetests --nologcapture --with-pylons=test.ini

To run the tests and produce a coverage report, first make sure you have coverage installed in your virtualenv (pip install coverage) then run:

nosetests --nologcapture --with-pylons=test.ini --with-coverage --cover-package=ckanext.xloader --cover-inclusive --cover-erase --cover-tests

Releasing a New Version of XLoader

XLoader is available on PyPI as https://pypi.org/project/ckanext-xloader.

To publish a new version to PyPI follow these steps:

1. Update the version number in the setup.py file. See PEP 440 for how to choose version numbers.

2. Update the CHANGELOG.

3. Make sure you have the latest version of necessary packages:

   pip install --upgrade setuptools wheel twine
   

4. Create source and binary distributions of the new version:

   python setup.py sdist bdist_wheel && twine check dist/*
   

   Fix any errors you get.

5. Upload the source distribution to PyPI:

   twine upload dist/*
   

6. Commit any outstanding changes:

   git commit -a
   git push
   

7. Tag the new release of the project on GitHub with the version number from the setup.py file. For example if the version number in setup.py is 0.0.1 then do:

   git tag 0.0.1
   git push --tags