Hi all,

I am writing to ask the community whether it is OK to deprecate
TreeMapMetaStore, as well as all the in-memory metastore manager,
metastore manager factory, and persistence types that rely solely on
TreeMapMetaStore.

As we know, these components are test-grade only, and not suitable for
production. They trigger a production readiness alert on Polaris
startup. It's a considerable amount of code that is virtually dead in
production.

It's also confusing for developers. E.g. the "transactional" metastore
is not transactional in the JDBC sense of the term, and thus not used
by JDBC persistence. It also has its quirks: some return statuses are
only returned by that manager.

However, these components are used in tests, and I agree that it's
useful to have an in-memory version of the persistence layer for
tests. But we have today two alternatives that are imho superior for
tests in polaris-runtime-service:

- JDBC persistence with H2 backend. There are already a few tests
using this setup.
- NoSQL persistence with InMemory backend.

Both alternatives test real production-grade persistence code.

And finally, TreeMapMetaStore is currently the default runtime
persistence in application.properties; and the Helm chart also
advertises it as the default. These are not sane defaults, imho. It's
always tricky to provide a good default for datastores, but since JDBC
persistence is included by default in the server image, I think that
including the H2 driver by default could give us a saner default while
keeping the out-of-the-box experience intact (the license is Category
A).

Concretely:

On the MetaStoreManagerFactory hierarchy, the following
implementations are completely in-memory:

- InMemoryPolarisMetaStoreManagerFactory: could be removed

- InMemoryAtomicOperationMetaStoreManagerFactory: could be removed

- LocalPolarisMetaStoreManagerFactory: is the base class of the two
above; imho it can be removed, but since it's an abstract class, it
may have been extended outside Polaris. But neither JDBC nor NoSQL use
it.

On the PolarisMetaStoreManager hierarchy:

- TransactionalMetaStoreManagerImpl: is only used by
LocalPolarisMetaStoreManagerFactory. JDBC and NoSQL do not use it.
Could be removed.

- TransactionWorkspaceMetaStoreManager however is a different beast,
in spite of the similar name. It is in use today on the commit path,
so should not be removed.

On the BasePersistence hierarchy:

- TreeMapTransactionalPersistenceImpl and its TreeMapMetaStore are
only used by InMemoryPolarisMetaStoreManagerFactory, and could be
removed;

- TransactionalPersistence and AbstractTransactionalPersistence: these
are supertypes of TreeMapTransactionalPersistenceImpl and thus only
used for in-memory. They imo can be removed, but they might have been
extended or implemented outside Polaris.

I think it's important to keep Polaris code tidy by removing unused,
unimplementable, or test-grade only components.

What are your thoughts?

Thanks,
Alex

Reply via email to