Re: [HACKERS] time stops within transaction

[Hiroshi Inoue]

Vadim Mikheev wrote: > > I am inclined to think that we should do SetQuerySnapshot in the outer > > loop of pg_exec_query_string, just before calling > > pg_analyze_and_rewrite. This would ensure that parse/plan accesses to > ^^^^^^^^^^^^^^ > Actually not - snapshot is passed as parameter to heap_beginscan... > And currently SnapshotNow is used everywhere. > I sometimes mentioned anxieties about the use of SnapshotNow, though I 've had no reasonable solution for it. SnapshotNow isn't a real snapshot and so it wouldn't be able to give us a complete consistency e.g. in the case "DDL statements in transaction block". However I couldn't think of any reasnoable way how to handle the following cases. We would have PREPARE statements in the near future. How does PREPARE use the same snapshot as the execution ? We would never be able to have shared catalog cache. We coulnd't delete dropped table files immediately after commit. ... Regards. Hiroshi Inoue