PostgreSQL 9.1 Open Items

Project Planning

See the PostgreSQL 9.1 Development Plan.

Meta-Issues

• standard_conforming_strings — readiness of drivers and applications
• Review "Long Term" list of items from PostgreSQL 9.0 Open Items

Open Issues

Not Blockers for Next Beta / RC

• Add regression tests for DDL commands with SSI

Resolved Issues

Issues Resolved Prior to RC1

• pg_upgrade fails if source database contains orphaned temp tables
• throw an error if you try to start from incomplete backup taken with pg_basebackup
  • proposed patch (seems unsuitable to put into 9.1 at this point, though)
• pg_restore fails with standard_conforming_strings
• Do latches have memory-ordering problems?
  • Yes they do ... but is 9.1 really vulnerable?
  • Careful code inspection shows we're safe. For now.
• SSI 2PC coverage test
  • associated bug fixed; need to shrink the test a bit before adding it

Issues Resolved Prior to Beta3

• plpython isn't connected up to tarball building
• ill-conceived grammar for NOT VALID constraints
• add a proof to README-SSI for the validity of the READ ONLY optimizations
  • Dan offered a concise proof, but we need a cite for one premise it uses
  • Kevin offered a self-contained proof, but it's much longer than Dan's
• FOREIGN TABLE doc fix
• SxactGlobalXmin computed incorrectly after abort
  • patch
• add a couple performance notes to the SSI docs
  • patch to do this
• Missing predicate locking in lossy bitmap heap scans
  • patch to fix
• btree_gist breaks some behaviors involving <> operators
• ALTER TABLE lock strength reduction patch is unsafe
• SSI atomic commit patch

Issues Resolved Prior to Beta2

• ALTER TYPE DROP + composite-typed col vs. pg_upgrade
• Memory leak in foreign scans
• vacuum sometimes fails to update relpages/reltuples
• SSI HOT chain traversal issue
• CTEs attached to data-modifying statements fail with rules (bug 6051)
• NOT VALID constraints don't dump properly
• HS slaves do not handle unlogged tables nicely (bug 6041)
• Make DDL commands SSI-aware
  • patch to fix (Problems found with patch, discussion and modifications in progress.)
• Domains over arrays no longer match ANYARRAY
• contrib/citext exposes implementation shortcomings in collation code (bug 6053)
• error in information_schema.element_types view definition (bug 5926)

Issues Resolved Prior to Beta1

• wal_buffers = -1 causes spurious chatter on reload
• is there a safeguard to prevent recovery from pausing before consistency is reached?
• Walreceiver crashes in AIX
  • Steve Singer can't reproduce, suggests a possible way this could be pilot error
• bug of the hot standby feedback
• GIST rewrite vs. pg_upgrade
• replication/README needs to be updated
• backend wrongly waits for sync rep even though max_wal_senders = 0
• reload of the configuration file should not cause the server to end unexpectedly
• Fix ecpg preprocessor regression
• Once sync_standbys_defined becomes true, there's no way for it to ever become false again. That can't be right.
• backend no longer needs to wait for replication when synchronous_standby_names is set to and configuration file is reloaded.
• pause_at_recovery_target needs to be added in recovery.conf.sample
• BUG #5842: Memory leak in PL/Python when taking slices of results
• hot standby feedback message needs to be explained at protocol-replication.html
• consider renaming ident to peer authentication on local connections
• serious problem by multiple backups
• FK constraints "NOT VALID" by default
• sync rep very slow when fsync=off on standby
• No longer need to check synchronous_standby_names and max_wal_senders at startup
• pg_last_xact_replay_timestamp limitations
  • rhaas says: It's not clear there's an action item here, so moving to resolved. Feel free to move back with more details.
• either remove write_location completely, or revert the change that broke it
  • original report of problem with write_location (but the other issues in that email are now fixed)
  • possible patch
• DO blocks leak memory
• replication timeout
• raise protocol version number
  • not done, would break things
• fix attinhcount tracking
• SSI: SIReadLock lines in pg_locks don't show pid
  • questions have been raised whether pid should be suppressed after connection closes or even maybe after transaction completion
• synchronous_commit and synchronous_replication
• conversion from integer literals to money type
• comments on SQL/MED objects
  • rhaas says: proposed patch
  • rhaas says: thom brown points out that I forgot USER MAPPINGs
  • fixing user mappings opens unduly large can of worms
• SSI: assertion failure on marking conflict-in due to race condition
  • existing patch to recheck after trading shared LW locks for exclusive should fix
• keywords table needs updating
• SSI: clumsy error handling results in generic message rather than more specific message with hint
• SSI: disable optimization when in subtransaction
• SSI: LOG message about SLRU wrap-around
• sync rep and fast shutdown
  • rhaas says: no easy resolution, i guess we'll leave this alone for now?
• SSI: three different HTABs contend for shared memory in a free-for-all
  • Heikki proposes "We'll need to teach dynahash not to allocate any more entries after the preallocation. A new HASH_NO_GROW flag to hash_create() seems like a suitable interface."
  • Alternatively, we could just use an initial allocation which matches the maximum entries, to ensure that all HTABs can allocate at least the configured maximum. (There's an existing patch for that.)
  • I committed a patch to add the new flag to hash_create() - Heikki
• SSI: failure to clean up some SLRU-summarized locks
  • existing patch to properly set commitSeqNo on the offending locks should fix
• ALTER TABLE ADD COLUMN not creating TOAST tables for inheritance children
• typed table DDL loose ends
• assorted collation issues
  • almost there...
• pl/python traceback fix
• gin indexes don't get used unless you vacuum
• SSI: High-contention UPDATE load on RAM disk database slows non-serializable transactions by a fraction of a percent
  • Or speeds it up by a fraction of a percent. Either way it's too close to tease out from the noise easily.
  • As insurance against any performance hit in extreme high contention loads, it might be worthwhile to do a quick return (without taking any locks) if no serializable transactions are active.
• SSI: UPDATE setting a TOASTed value is broken, regardless of transaction isolation level
  • proposed patch
• foreign table permissions issues
• CREATE TABLE IF NOT EXISTS doesn't work
• Typed-tables patch broke pg_upgrade
  • proposed patch (tt2v2-binary-upgrade.patch)