# Conflicts:
#   swh/storage/tests/test_cli_cassandra.py

This makes swh.journal properly optional when using only the pytest_plugin.

This should allow us to avoid having many modules (test-)depend on
`swh-storage[testing]`.

Cassandra keyspace is already available in the keyspace attribute of
the swh_storage fixture.

This prevents new connections to the cassandra cluster as a side effect
that can make tests flaky when executed in parallel.

This is now completely handled as a journal client in swh.counters, so
this code has not run in years (except in tests).

It also introduces a strict dependency between swh.storage and
swh.journal via swh.counters, which should not be needed.

It allows to avoid tests flakiness by preventing the creation of a new storage
instance when tearing down the swh_storage_cassandra_backend_config fixure.

Storage content cannot be persisted across multiple memory storage
instances in the same test so some tests were failing.

Some cassandra request timeouts have been observed on Jenkins
when tearing down the swh_storage_cassandra_backend_config
fixture so execute those requests with retries to mitigate
these issues.

This will need coordination with other MR migrating other part of swh to
psycopg3, more notably, swh-core.

Updates various imports and names.

Adjusts syntax query and parameter to what psycopg3 expect. The `origins_are_blocked` of the Blocking proxy class needed significant rework for equivalent features.

Dropped timezone related work around that existed for psycopg2 < 2.9.0

Adjust the way we retrieve the connection information. This update as the same caveas as for swh-core and needs some double check.

Use the dedicated `psycopg_pool` package for connection pull.

Adds a small `execute_values_generator` helper to match existing code
expectation better (reducing the complexity of this diff).

Drops a call to the deprecated and noop `app.setup_psycopg_errorhandlers`

Add more type to help mypy and catch errors.

Modify swh_storage_cassandra_cluster to ensure a single instance of
cassandra is spawned when running tests in parallel with pytest-xdist.

Explicitly shutdown clusters create by CqlRunner instances to avoid the
display of traceback related to Python threads when cassandra related
fixtures are tear downed.

These changes significantly reduce the time to execute a tox session
(more than a x2 speedup).

It allows to avoid the display of tracebacks related to Python threads
when cassandra related pytest fixtures are tear downed.

Side effect of adding typing to swh.core.utils.grouper.

This notably removes all the warnings related to comparing a
ModelObjectType enum with a string literal.

As a consequence ModelObjectType enum are used as dict keys
instead of literal strings.

vlorentz authored 1 month ago and vlorentz committed 1 month ago

https://gitlab.softwareheritage.org/swh/devel/swh-loader-metadata/-/blob/c2cdfd5c86d2783ce28c80f84df4930c2eba3cfc/swh/loader/metadata/gitea.py#L121-138

This is used by recent versions of the cassandra Debian package to
override the log directory.

Antoine Lambert authored 1 month ago and Antoine Lambert committed 1 month ago

This was used at the time we were building debian packages for
swh components but we no longer do that.

hypothesis uncovered some edge cases datetime objects where Python will
produce a timestamp offset by one second due to an unexpected rounding,
likely due to float point precision issue during computation.

Set datetime microseconds to 0 before calling the timestamp method to
avoid such side effect and ensure converted date will be the same as
the input one.

Fixes #4720.

Bump development tools: mypy, codespell, isort, ...

Move all tools configuration in pyproject.toml.

Remove no longer needed mypy overrides.

It expected a simple cls cassandra so far but it does not reflect the
production use case (we deploy storage cassandra with pipeline configuration).

It has been observed calling that method can lead to cassandra
query timeout when a snapshot contains a large amount of branches.

The performance bottleneck seems to be the use of a custom cassandra
aggregate function computing the number of branches per target type.

Removing the use of that function and performing the aggregation
client side significantly improves query performance and makes
the timeout issue goes away.

Fixes #4719.

The assertion was failing when executing the code so use a different
one to keep mypy happy.

Also add test to ensure the diagram generation code is covered.

This should speed up a bit tests execution on Jenkins.

Use pytest option --dist=loadfile to group tests by their containing file
in order to avoid spawning too many cassandra or postgres instances.

Recent changes in swh-storage added the PartitionsManagementInterface
but this interface is only implemented on real storage backends, not
proxies.

Any client code overriding the swh_storage_backend_config to return a
storage pipeline (loaders for instance) ends up with error otherwise.

Through a `table_option` dict in the config, mapping table names to what to add to the `WITH` clause of the `CREATE TABLE` clause.

This abstracts this interface from the PostgreSQL-specific implementation,
allowing the object_references table to be sharded in Cassandra.