pgdogdev/pgdog v0.1.64 on GitHub

Performance

Two-phase commit in auto mode would not commit correctly with multi-tuple INSERT statements
Empty query (;) used by Rails for health checking a connection would be blocked inside a failed transaction
Query cancellation would leak between connections if the query finished on its own before the cancellation request was processed by Postgres
Primary key discovery wouldn't work for read-only users which would cause automatic primary key injection to break in sharded deployments if the read user was used to fetch the database schema (race condition)
UNLISTEN * wouldn't do anything, it now correctly unsubscribes the client from all channel notifications
Fast, consecutive RELOADs could propagate "pool is offline" to connected clients
Replica lag formula would break if user changed the timezone session parameter; this is because the replica lag monitor re-uses connections from the user pool & and we don't (intentionally) reset parameters on connection checkin. All time functions in the replica lag query now use AS TIME ZONE 'UTC'
read_write_split = "exclude_primary" was allowing reads on primary if no replicas were available
Inject Describe into extended protocol used in cross-shard queries and use the returned RowDescription to decode rows when performing cross-shard aggregation; this avoids data type mismatches during resharding when we automatically upgrade integer columns to bigint

Correctly handle count(*)::numeric type cast in cross-shard SELECT
Filter replicas from the load balancer using their last known LSN by providing the LSN as part of a query comment, e.g., /* pgdog_lsn: 0/0 */ SELECT [...]
Add read_write_split setting to users in users.toml
Partial support for HAVING clause in aggregates
Support sorting/aggregating on inet, money and cidr data types

Added snakeoil certs back into the Docker image. Don't use them in prod, but they are useful for testing. Security scanners flag those; just ignore it.

Full Changelog: v0.1.63...v0.1.64