invenio-rdm-migrator 1.0.0a7

InvenioRDM module for data migration.

DataCite-based data model for Invenio.

Development

Install

Make sure that you have libpq-dev installed in your system. See psycopg installation instructions for more information.

Choose a version of search and database, then run:

pip install -e .

Tests

./run-tests.sh

How to run it

To run the migration you need:

• A running InvenioRDM instance.

• If your data contains references to other records (e.g. vocabularies), then it is also required to run the setup step.

invenio-cli services setup --force --no-demo-data

• Install Invenio-RDM-Migrator, any other dependencies must be handled in the Pipfile of your instance.

$ pip install invenio-rdm-migrator

• Create/edit the configuration file on your instance, for example streams.yaml:

records:
    extract:
        filename: /path/to/records.json
    load:
        db_uri: postgresql+psycopg2://inveniordm:inveniordm@localhost:5432/inveniordm
        tmp_dir: /tmp/migrator

• You will need to create a small python script putting together the different blocks of the ETL. You can find an eample at my-site/site/my_site/migrator/__main__.py.

from invenio_rdm_migrator.streams import StreamDefinition
from invenio_rdm_migrator.streams.records import RDMRecordCopyLoad

if __name__ == "__main__":
    RecordStreamDefinition = StreamDefinition(
        name="records",
        extract_cls=JSONLExtract,
        transform_cls=ZenodoToRDMRecordTransform,
        load_cls=RDMRecordCopyLoad,
    )

    runner = Runner(
        stream_definitions=[
            RecordStreamDefinition,
        ],
        config_filepath="path/to/your/streams.yaml",
    )
    runner.run()

• Finally, you can execute the above code. Since it is in the __main__ file of the python package, you can run it as a module:

$ python -m my_site.migrator

• Once the migration has completed, in your instance you can reindex the data. For example, for users and records it would look like:

$ invenio-cli pyshell

In [1]: from invenio_access.permissions import system_identity
In [2]: from invenio_rdm_records.proxies import current_rdm_records_service
In [3]: from invenio_users_resources.proxies import current_users_service

In [4]: current_users_service.rebuild_index(identity=system_identity)
In [5]: current_rdm_records_service.rebuild_index(identity=system_identity)

Implement your {Extract/Transform/Load}

There are for packages in this module extract, transform, load, and streams. The first three correspond to the three steps of an ETL process. The streams package contains the logic to run the process and different stream-specific implementations of ETL classes (e.g. records).

Extract

The extract is the first part of the data processing stream. It’s functionality is quite simple: return an iterator (e.g. of records), where each yielded value is a dictionary. Note that the data in this step is _transformed_ to an extent, but only in format (e.g. JSON, XML), not in content. For example, to implement a XMLExtract class:

class XMLExtract(Extract):
...

    def run(self):
        with open("file.xml") as file:
            for entry in file:
                yield xml.loads(entry)

Transform

The transformer is in charge of modifying the content to suit, in this case, the InvenioRDM data model (e.g. for records) so it can be imported in the DB. It will loop through the entries (i.e. the iterator returned by the extract class), transform and yield (e.g. the record). Diving more in the example of a record:

To transform something to an RDM record, you need to implement streams/records/transform.py:RDMRecordTransform. For each record it will yield what is considered a semantically “full” record: the record itself, its parent, its draft in case it exists and the files related them.

{
    "record": self._record(entry),
    "draft": self._draft(entry),
    "parent": self._parent(entry),
}

This means that you will need to implement the functions for each key. Note that, only _record and _parent should return content, the others can return None. In this case we will need to re-think which methods should be abstractmethod and which ones be defaulted to None/{}/some other default in the base. You can find an example implementation at zenodo-rdm/site/zenodo_rdm/migrator/transform.py:ZenodoToRDMRecordTransform.

Some of these functions can themselves use a transform/base:Entry transformer. An _entry_ transformer is a one layer deeper abstraction, to provide an interface with the methods needed to generate valid data for part of the Transform class. In the record example, you can implement transform.base:RDMRecordEntry, which can be used in the RDMRecordTransform._record function mentioned in the code snippet above. Note that implementing this interface will produce valid _data_ for a record. However, the _metadata_ is not interfaced. It is an open question how much we should define these interfaces and avoid duplicating already existing Marshmallow schemas.

At this point you might be wondering “Why not Marshmallow then?”. The answer is “separation of responsibilities, performance and simplicity”. The later lays with the fact that most of the data transformation is custom, so we would end up with a schema full of Method fields, which does not differ much from what we have but would have an impact on performance (Marshmallow is slow…). Regarding the responsibilities part, validating - mostly referential, like vocabularies - can only be done on _load_ where RDM instance knowledge/appctx is available.

Note that no validation (not even structural) is done (at the moment) in this step.

Load

The final step to have the records available in the RDM instance is to load them. The available load/postgresql:PostgreSQLCopyLoad will carry out 2 steps:

• 1. Prepare the data, writing one DB row per line in a csv file:

$ /path/to/data/tables1668697280.943311
    |
    | - pidstore_pid.csv
    | - rdm_parents_metadata.csv
    | - rdm_records_metadata.csv
    | - rdm_versions_state.csv

2. Perform the actual loading, using COPY. Inserting all rows at once is more efficient than performing one INSERT per row.

Internally what is happening is that the prepare function makes use of TableGenerator implementations and then yields the list of csv files. So the load only iterates through the filenames, not the actual entries.

A TableGenerator will, for each value in the received iterator, yield one or more rows (lines to be written to the a csv file). For example for a record it will yield: recid, DOI and OAI (PersistentIdentifiers), record and parent metadata, etc.

Notes

Shared cache between streams

During a migration run, there is a need to share information across different streams, e.g populate communities before records and keep the map between community slug names and autogenerated community ids, or at the same stream across different TableGenerator instances, e.g on records stream we keep the “seen” parent ids so we can update the information of the parent for different record versions. For that reason, we pass a cache dict, that can change in the future in a type of persistent storage e.g redis, in each stream.load step so streams can populate/consume the cache. The cache for each stream can also be populated in each stream configuration like below in your streams.yaml:

records:
    extract:
        filename: /path/to/records.json
    load:
        cache:
            communities:
                community_slug: <community_id>
        db_uri: postgresql+psycopg2://inveniordm:inveniordm@localhost:5432/inveniordm
        tmp_dir: /tmp/migrator

When the runner will instantiate each stream will merge the existing state of the cache with whatever is provided in the stream configuration. That means, that the stream configuration takes precedence and can override the whole cache before the stream runs! Any cache state that exists before is overridden for the rest of the migration run.

Infrastructure

While now we are chaining the iterator from one step into the other in the streams, the idea is that all three steps will pull/push to/from queues so they can be deployed in different parts of the system (e.g. the load part in the worker nodes).

Others

• Using generators instead of lists, allows us to iterate through the data only once and perform the E-T-L steps on them. Instead of loop for E, loop for T, loop for L. In addition, this allows us to have the csv files open during the writing and closing them at the end (open/close is an expensive op when done 3M times).

Changes

Version 1.0.0

• Initial public release.