PostgreSQL 9.2 Open Items

Project Planning

See the PostgreSQL 9.2 Development Plan.

Meta-Issues

• Review "Long Term" list of items from PostgreSQL 9.0 Open Items and PostgreSQL 9.1 Open Items

Blockers for 9.2

(none known)

Not Blockers for 9.2

• Put a bound on memory usage when building large GiST indexes. (see last paragraph of this mail)
  • seems like this is a nice to have, not essential -rhaas
• simple_prompt uses wrong charset (Windows)
  • is this bug any worse than any other open bug? we don't usually hold up releases for those. -rhaas
  • it's not new in 9.2, either - tgl
• missing replication admin functions
  • more discussion and engineering is needed here, and the time for forcing initdb is past -rhaas
• Should we fix tuple limit handling, or redefine 9.x behavior as correct?
  • The consensus seems to be to change the documentation to match the current behavior, so this is just a pre-existing documentation bug, not a blocker.
• Recheck performance claims for checkpointer process split?
  • The failure-to-fsync bug in the original coding makes the previous measurements entirely untrustworthy. However, that change wasn't sold on pure performance grounds, so we'd probably keep it even if there is some adverse impact.
• pg_ctl's adjust_data_dir fails on Windows when -D points to config-only directory
  • proposed patch
  • This is not a regression because pg_ctl never really worked before anyway when -D points to a config-only directory.

Resolved Issues

• streaming backup should not be considered as sync standby.
• Documentation needs updating wrt visibility map/index-only scans
• change GiST implementation for range types?
• pglesslog still referenced in docs, but no 9.1 support
• Should we change the behavior of extract(epoch from timestamp)?
• search_path behavioral change
• Replace bytea_agg with string_agg(bytea,bytea)?
• pg_size_pretty(numeric)
• Reconsider the names of valid values of synchronous_commit
  • call for opinions
• nodes/*funcs.c inconsistencies
• Parameterized-path costing still needs some work
• Change syntax for non-inherited CHECK constraints?
• Aggressive memory consumption in {ts,array}_typanalyze
• Auto-ANALYZE can prevent all-visible flags from getting set, leading to failure to use index-only scans
  • I've hacked the regression test case to prevent the intermittent failure. Fixing auto-ANALYZE so that the problem doesn't arise would be nice, but it seems like a research project; added that to TODO.
• patch: plpython triggers are broken for composite-type columns
• Inconsistent units in pg_stat_statements
• psql omits row count under "\x auto"
• Update the sample ps output in "monitoring-ps" section of the manual to include the new checkpointer process
• Release-note PL case folding incompatibility
• checkpointer gets confused after crash-and-restart
  • On closer inspection, this was pilot error, namely misuse of the postmaster -T feature.
• Checkpointer latch loop [Simon]
  • I've hacked that area of the code until it looks OK to me, is there something else you had in mind? -- tgl - no thats fine - sr
• Double check bgwriter/checkpointer comments
  • I made a pass over these, does anyone want to do more? -- tgl
• Check new processes listed correctly in postmaster comments
  • I made a pass over these, does anyone want to do more? -- tgl
• Make XLOG_HEAP2_VISIBLE records generate recovery snapshot conflicts so that IndexOnlyScans work on Hot Standby + bug fix by sr
  • analysis
  • committed patch - still would be nice to throw a soft conflict instead of a hard one
• Correct docs for SP-GIST so that example implementations shown/referred to/examples given - unusably opaque at present
  • Huh? There is a pointer to the example opclasses, and in general the SPGiST docs seem to me to be at least as good as those for GiST or GIN - tgl
• pg_basebackup -x stream from the standby gets stuck if there is no traffic in the database
• pg_receivexlog requires manually restarting whenever the sending server restarts
  • A brief discussion offline came up with the suggested options --auto-reconnect --auto-disconnect
• Fix bug about not using InHotStandby when throwing recovery conflicts [Simon]
• Checkpointer doesn't send stats promptly [Simon]
• Fix start order: check pointer must start before bgwriter to catch fsync requests [Simon]
• Bgwriter needs to close file handles [Simon]
• B-tree page deletion boundary cases [Simon]
• libpq URL syntax needs work for unix-socket addresses
  • proposed patch, commited
• pg_size_pretty(numeric)
  • decided not to change this
• Start warning on CREATE {GLOBAL | LOCAL} TEMPORARY TABLE? [Simon]
• incorrect handling of the timeout in pg_receivexlog
• warning on incorrect parameters in libpq URL format is a potential security problem and should be changed to error
• scary checkpoint stuff
• pg_basebackup/pg_receivexlog vs sync rep
• alter table drop constraint does not work on inherited master table
• Reconsider CHECK NO INHERIT syntax?
• Should we remove libpq row processor callback API in favor of row-at-a-time resultsets?
• SP-GiST is unsafe for use by hot standby slaves
• "pg_basebackup --xlog-method stream" may fail to take backup silently
• WAL files which were restored from the archive are archived again. thread continuation
  • proposed patch [committed by Simon]
• keepalives
• View options are problematic for pg_dump
• Checkpointer process split broke fsync'ing
• pl/perl crash on win32 - pl/perl crashes on win32/MSVC++ builds with ActiveState Perl (the only distribution that we can use with MSVC++).
• Cascading replication and recovery_target_timeline='latest'

Long-term Issues

• GiST indexes vs fuzzy comparisons used by geometric types
  • Alexander proposed a narrow patch that would support the current behavior, but I don't have a lot of confidence in it - tgl
  • The larger issue is whether we should modify or get rid of the whole fuzzy-comparison behavior in these data types.