Re: Logical decoding on standby

On 22 November 2016 at 03:14, Andres Freund  wrote:
> Hi,
>
> On 2016-11-21 16:17:58 +0800, Craig Ringer wrote:
>> I've prepared a working initial, somewhat raw implementation for
>> logical decoding on physical standbys.
>
> Please attach. Otherwise in a year or two it'll be impossible to look
> this up.

Fair enough. Attached. Easy to apply with "git am".

I'm currently looking at making detection of replay conflict with a
slot work by separating the current catalog_xmin into two effective
parts - the catalog_xmin currently needed by any known slots
(ProcArray->replication_slot_catalog_xmin, as now), and the oldest
actually valid catalog_xmin where we know we haven't removed anything
yet.

That'll be recorded in a new CheckPoint.oldestCatalogXid field and in
ShmemVariableCache ( i.e. VariableCacheData.oldestCatalogXid ).

Vacuum will be responsible for advancing
VariableCacheData.oldestCatalogXid by writing an expanded
xl_heap_cleanup_info record with a new latestRemovedCatalogXid field
and then advancing the value in the ShmemVariableCache. Vacuum will
only remove rows of catalog or user-catalog tables that are older than
VariableCacheData.oldestCatalogXid.

This allows recovery on a standby to tell, based on the last
checkpoint + any xl_heap_cleanup_info records used to maintain
ShmemVariableCache, whether the upstream has removed catalog or
user-catalog records it needs. We can signal walsenders with running
xacts to terminate if their xmin passes the threshold, and when they
start a new xact they can check to see if they're still valid and bail
out.

xl_heap_cleanup_info isn't emitted much, but if adding a field there
is a problem we can instead add an additional xlog buffer that's only
appended when wal_level = logical.

Doing things this way avoids:

* the need for the standby to be able to tell at redo time whether a
RelFileNode is for a catalog or user-catalog relation without access
to relcache; or
* the need to add info on whether a catalog or user-catalog is being
affected to any xlog record that can cause a snapshot conflict on
standby; or
* a completely reliably way to ensure hot_standby_feedback can never
cease to affect the master even if the user does something dumb

at the cost of sometimes somewhat delaying removal of catalog or
user-catalog tuples when wal_level >= hot_standby, a new CheckPoint
field, and a new field in xl_heap_cleanup_info .

The above is not incorporated in the attached patch series, see the
prior post for status of the attached patches.

-- 
 Craig Ringer                   http://www.2ndQuadrant.com/
 PostgreSQL Development, 24x7 Support, Training & Services