v0.1.38

What's Changed

• feat: allow using SCRAM hash in users.toml as password_hash, if using server_auth = "rds_iam"; no passwords stored in any pgdog config file anymore
• feat: add retries for COPY_DATA command (used in resharding); allows to retry long-running COPY operations instead of failing @meskill
• feat: server_auth with Microsoft Workload Identity @MagicAbdel
• feat: parse query comments using custom, faster parser; speed up the parser by 50%, 15% TPS improvement across the board when using simple protocol (e.g. Rails with prepared_statements: false)
• fix: support servers that don't allow query cancellation (e.g., chaining RDS Proxy and PgDog now works)
• fix: crashes in edge cases when query_parser_engine = "pg_query_raw"
• feat: add max_pool_size, alias for default_pool_size (and pool_size)
• fix: re-initialize pub/sub channel after failover
• fix: add proper handling for pool restarts during resharding @meskill

New Contributors

• @MagicAbdel made their first contribution in #906

Full Changelog: v0.1.37...v0.1.38

v0.1.37

What's Changed

• fix: add bounds checks to some buffer accesses by @jaggederest
• fix: ProtocolOutOfSync error
• fix: count(*) in query broke CTE write check
• fix: return ProtocolViolation instead of syntax error in admin DB; makes tools like pgcli work out of the box
• feat: support multiple user/server passwords; supports zero-downtime password rotation
• feat: handle SET in single-primary / only replicas configuration automatically
• feat: add regex parser trigger to minimize parsing when not needed
• feat: add parallel copy override for speeding up table copy during resharding
• feat: add log throttle, to suppress duplicative logs during error storms
• feat: detect advisory locks with regex
• feat: support arrays as aggregate targets in cross-shard queries @jaggederest

Full Changelog: v0.1.36...v0.1.37

v0.1.36

What's Changed

• feat: parallelize copy data & schema sync across shards
• feat: change default setting value for query_cache_limit to 1000 to avoid heavy memory usage on first deployment
• feat: log warning if re-sharding replication is missing rows; this if only impactful is using copy-data without --replication-only flag and shouldn't happen under normal state
• feat: add 53-bit unique ID generator for apps that pass IDs directly to JS frontends
• feat: add user-friendly interface to create sharded sequences (pgdog.install_sharded_sequence pl/pgsql function)
• fix: use correct backend data type for cross-sharded count(*), fixes count(*)::int returning incorrect count
• fix: oid is u32, not i32; affected old databases with schema churn (oid exceeded 2.2B)
• fix: omnisharded table re-sharding: copy and replication copied the same table N times (N = number original shards)
• chore: regression testing for unique id
• chore: migrate to RWX for CI by @kylekthompson
• chore: resharding integration tests
• fix: detect SELECT ... FOR UPDATE inside CTEs
• fix: potentially dropping child table rows during streaming replication / re-sharding

New Contributors

• @kylekthompson made their first contribution in #862

Full Changelog: v0.1.35...v0.1.36

v0.1.35

What's Changed

• feat: DDL-only/DML-only mirroring
• fix: much faster schema loading query, works with databases with many tables
• fix: invalid keep-alive settings are just a warning now, won't cause a crash
• fix: race in extended_anonymous prepared statements
• fix: deadlock in sharded SELECT ... WHERE IN ($1, $2, $3)
• fix: handle LIMIT $1 where $1 is null using the extended protocol
• fix: logical replication Delete message generated incorrect parameters with compound identity columns
• chore: add postgres.js tests to CI

Full Changelog: v0.1.34...v0.1.35

v0.1.34

In-transaction connections deadlock

If you're using v0.1.31 or later, please upgrade to v0.1.34 asap. This release fixes a deadlock caused by clients disconnecting mid-transaction.

What's Changed

• fix: client disconnect inside transaction caused deadlock
• chore: fix clippy issues by @meskill

Full Changelog: v0.1.33...v0.1.34

v0.1.33

What's Changed

• feat: PgDog-supported plugin for routing queries using table names (e.g., for primary-only routing)
• feat: allow passthrough auth user to change its password
• fix: Passthough -> Passthrough typo @rissson
• fix: mirror prepared statement mode in session pooling @costi
• fix: correctly handle Relation message sent inside transaction during logical replication (used for resharding)
• fix: correctly forward only one Postgres error to the client, using the error code provided by Postgres server
• fix: correctly handle FATAL / PANIC errors (force close the connection in the pooler) without banning host
• fix: dependabot reporting security issues in integration tests repo-wide (e.g., Rails, Python, etc.) @meskill

New Contributors

• @rissson made their first contribution in #829
• @costi made their first contribution in #827

Full Changelog: v0.1.32...v0.1.33

v0.1.32

What's Changed

• feat: weighted load balancing
• fix: schema sharding in session mode with cross_shard_disabled
• feat: add configurable log format and level by @levish0

New Contributors

• @levish0 made their first contribution in #822

Full Changelog: v0.1.31...v0.1.32

v0.1.31

What's Changed

• feat: online resharding with zero downtime, incl. admin commands to orchestrate it
• feat: RDS IAM authentication support for server connections @Tolsto
• feat: handle RESET command
• feat: ban replicas if replica lag is high (configurable)
• feat: config linting and validation with jsonschema, incl. language server support @meskill
• fix: replication stream broke due to parser changes, added regression tests
• fix: update verify_ca implementation for rustls 0.23, fixes Aurora TLS connections with verify_ca (verify_full worked already) @quentindemetz
• fix: connection refused to PgDog if a database is down on proxy startup
• fix: binary COPY crash with failed to fill whole buffer, caused by incorrect buffering of partial packets
• fix: segfault in pg_query deparse_raw caused by empty C strings
• fix: LISTEN/NOTIFY/UNLISTEN were blocked in session mode if pub_sub_channel_size was set to 0 by @mnbbrown
• fix: transactions were incorrectly buffered in session mode and swallowed in transaction mode if advisory locks were used
• fix: sending partial requests to backend after abrupt client disconnect, causing server connections to be stuck in ClientRead and active state
• fix: exclude_primary should read from primary if there are no replicas @mijoharas
• fix: add more params to the ignore list, like is_superuser, which caused errors when clients modified them with SET
• chore: use OUT_DIR for autogenerated files, add bindings.rs to .gitignore @mijoharas

New Contributors

• @quentindemetz made their first contribution in #786
• @mnbbrown made their first contribution in #800
• @Tolsto made their first contribution in #793

Full Changelog: v0.1.30...v0.1.31

v0.1.30

What's Changed

BREAKING: user settings take priority

User settings in users.toml now take priority over database settings in pgdog.toml. This was the opposite previously. For example:

users.toml

[[users]]
name = "pgdog"
database = "prod"
pool_size = 20

pgdog.toml

[general]
default_pool_size = 10

[[databases]]
name = "prod"
host = "127.0.0.1"
pool_size = 25

The connection pool for user pgdog will be 20.

🙏 @mijoharas

Changelog

• feat: support OFFSET in cross-shard SELECT queries
• feat: support for plugin API versioning @meskill
• fix: incorrect error handling when queries are pipelined (psycopg, Sequelize, etc.)
• feat: client_connection_recovery defaults to drop
• refactor: make schema serializable for Enterprise Edition integration
• feat: schema fetches foreign keys to support tables without sharding keys

New Contributors

• @meskill made their first contribution in #770

Full Changelog: v0.1.29...v0.1.30

v0.1.29

What's Changed

• feat: expose default_pool_size / pool_size, prepared_statements_limit and query_cache_limit as Prometheus metrics @Adi-Goll
• feat: support cross-shard LIMIT (but not OFFSET, yet)
• feat: support multiple SET statements in one query
• feat: add client_connection_recovery setting which allows to close client connections that have received a checkout timeout / all replicas down error from PgDog
• fix: race condition in Sequelize (Bind, Exexute, Flush, Sync) that caused record creation to be acknowledged before transaction committed in Postgres
• fix: don't attempt to recover server connections that have received partial client requests (hard to determine state of connection)
• fix: handle PortalSuspended message correctly
• chore: Prisma, Sequelize tests
• fix: place Sync into its own ClientRequest in spliced requests (Java driver)

New Contributors

• @Adi-Goll made their first contribution in #740

Full Changelog: v0.1.28...v0.1.29