> On 19 Apr 2017, at 12:37, Petr Jelinek <petr.jelinek@2ndquadrant.com> wrote:
>
> On 18/04/17 13:45, Stas Kelvich wrote:
>> Hi,
>>
>> currently logical replication worker uses ApplyContext to decode received data
>> and that context is never freed during transaction processing. Hence if publication
>> side is performing something like 10M row inserts in single transaction, then
>> subscription worker will occupy more than 10G of ram (and can be killed by OOM).
>>
>> Attached patch resets ApplyContext after each insert/update/delete/commit.
>> I’ve tried to reset context only on each 100/1000/10000 value of CommandCounter,
>> but didn’t spotted any measurable difference in speed.
>>
>> Problem spotted by Mikhail Shurutov.
>>
>
> Hmm we also do replication protocol handling inside the ApplyContext
> (LogicalRepApplyLoop), are you sure this change is safe in that regard?
Memory is bloated by logicalrep_read_* when palloc happens inside.
I’ve inserted context reset in ensure_transaction() which is only called in beginning
of hande_insert/update/delete/commit when data from previous call of that
function isn’t needed. So probably it is safe to clear memory there. At least
i don’t see any state that can be accessed later in this functions. Do you
have any specific concerns?
Stas Kelvich
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company
