Planning Notes

Brainstorming scratchpad. Takeaway: going with FDW APIs, using postgres_fdw's connection.c and deparse.c as inspiration.

Features

Comprehensive list of desired features for each version.

First Version

First internal release.

• Single-row inserts through master with replication
• Naïve (UNION ALL-style) SELECT queries through master (no JOIN, sub-SELECT, or common table expression support)
• Fall-through to optimized (CitusDB) logic for existing customers
• Shard pruning for absolute equality predicates in SELECTs
• Hash partitioning across a fixed number of partitions using a single column's value
• Single algorithm for task assignment (random or first-replica)
• Strict consistency across replicats (users must intervene if a node becomes unhealthy)
• One-shot schema propagation (up-front definition of tables)
• Seamless upgrade to CitusDB from PostgreSQL for extension users

Future Versions

• Failure modes
• Citus modifications
• UPDATE/DELETE
• Range partitions
• Schema modification
• Trans-table partition schemes

Trouble Spots

Possible problems.

• Collations
• Domains
• Custom types
• Inheritance
• Exclusion constraints (unless on hash key)
• Foreign keys (unless both tables use same hash scheme)
• Uniqueness constraints (unless on hash key)
• Oids
• Sequences
• Storage parameters

Takeaways

More iterative reviews; frequently/often.

Plan discrete, compilable, runnable chunks.