Thoughts on scale of Postgres where mooncake runs?

[umeshkacha]

Mooncake looks promising with great vision of giving best of OLTP and OLAP. Just curious when usually OLAP use case needs to store much larger data in TBs, how will Postgres scale in such a case? Are you by design hoping to expect users to keep data in blob storages like S3 and pull into Postgres via DuckDB S3 functions and write it back to S3?

[dpxcc]

Just curious when usually OLAP use case needs to store much larger data in TBs, how will Postgres scale in such a case?

I believe a single-node DuckDB can comfortably manage datasets ranging from gigabytes to several terabytes. For datasets beyond that, there are a couple of options for scaling:

1. Use citus to turn Postgres into a distributed database. Note that pg_mooncake and citus don't work well together right now and it's not an easy task because both extensions make significant changes to how Postgres plans and executes queries
2. Use external compute to scale beyond a single node. We may explore our cloud offering to address this scenario in the future

Are you by design hoping to expect users to keep data in blob storages like S3 and pull into Postgres via DuckDB S3 functions and write it back to S3?

We support storing data in the local filesystem. If mooncake.allow_local_tables is set to true and mooncake.default_bucket is unset, data will be written to the <PG_DATA_DIR>/mooncake_local_tables/mooncake_<database>_<table>_<oid>/ directory. Additionally, you can even mount <PG_DATA_DIR>/mooncake_local_tables/ to your in-house shared storage layer

However, on Neon serverless Postgres, mooncake.allow_local_tables is set to false, and you must store data in blob storage. This is because any data stored locally will be lost after a server restart