Thread: Performance penalty during logical postgres replication
Hi,
We are doing a logical postgres replication from Postgres 11 to 12. Our database is around 700GB (8 cpu's, 32 GB).
During the replication process, at some point, we see a huge performance penalty on a particular table. This table acts as a queue with lots of inserts and deletes happening throughout the day. For most of the time this table is empty, but during this performance penalty the number of rows in this table grows to 10.000 rows, and processing is not fast enough to empty this table. Main reason for this (as far as we see) is that the performance of the query for selecting the next row to process drops from < 10MS to 400MS. This eventually causes too much cpu load on the Primary and we have to cancel the replication process.
We already tried the initial load three times, and it consistently fails with the same "error". Last try was a per table approach and excluding this "queue" table.
After cancelling the replication the query is fast again and the load on the Primary goes back to normal. We see that this happens when replicating large tables (> millions of rows). During this performance penalty the explain of the query selecting the next row from this table tells us it is doing a sequential scan (there is an index but it is not used).
- What could cause this performance penalty?
- Is this something other people experienced as well during the initial load of a logical replication with large tables?
- We are now thinking of temporarily increasing the number of CPU's and RAM for the migration. Would this help in this case?
Thanks in advance,
Lars
On Wed, Dec 9, 2020 at 2:21 AM Lars Vonk <lars.vonk@gmail.com> wrote:
Hi,We are doing a logical postgres replication from Postgres 11 to 12. Our database is around 700GB (8 cpu's, 32 GB).During the replication process, at some point, we see a huge performance penalty on a particular table. This table acts as a queue with lots of inserts and deletes happening throughout the day. For most of the time this table is empty, but during this performance penalty the number of rows in this table grows to 10.000 rows, and processing is not fast enough to empty this table. Main reason for this (as far as we see) is that the performance of the query for selecting the next row to process drops from < 10MS to 400MS. This eventually causes too much cpu load on the Primary and we have to cancel the replication process.We already tried the initial load three times, and it consistently fails with the same "error". Last try was a per table approach and excluding this "queue" table.After cancelling the replication the query is fast again and the load on the Primary goes back to normal. We see that this happens when replicating large tables (> millions of rows). During this performance penalty the explain of the query selecting the next row from this table tells us it is doing a sequential scan (there is an index but it is not used).- What could cause this performance penalty?- Is this something other people experienced as well during the initial load of a logical replication with large tables?- We are now thinking of temporarily increasing the number of CPU's and RAM for the migration. Would this help in this case?Thanks in advance,Lars
I would guess that you are hitting bad plans (sequential scan instead of index) because auto-vacuum is not picking up this table for vacuum/analyze as often as needed. You could try a periodic 'vacuum analyze table' while the logical replication is going, or just look at your autovacuum settings and tune the system to be more aggressive with turning down scale factor and cost delay. If the table is primarily "insert only" particularly with this initial load of data, then the autovacuum may not pick it up at all. That pain point is mitigated with PG13's new behavior to kick off autovacuum/analyze based on inserts also.
ср, 9 дек. 2020 г. в 10:21, Lars Vonk <lars.vonk@gmail.com>:
We are doing a logical postgres replication from Postgres 11 to 12. Our database is around 700GB (8 cpu's, 32 GB).During the replication process, at some point, we see a huge performance penalty on a particular table. This table acts as a queue with lots of inserts and deletes happening throughout the day. For most of the time this table is empty, but during this performance penalty the number of rows in this table grows to 10.000 rows, and processing is not fast enough to empty this table. Main reason for this (as far as we see) is that the performance of the query for selecting the next row to process drops from < 10MS to 400MS. This eventually causes too much cpu load on the Primary and we have to cancel the replication process.
We already tried the initial load three times, and it consistently fails with the same "error". Last try was a per table approach and excluding this "queue" table.After cancelling the replication the query is fast again and the load on the Primary goes back to normal. We see that this happens when replicating large tables (> millions of rows). During this performance penalty the explain of the query selecting the next row from this table tells us it is doing a sequential scan (there is an index but it is not used).- What could cause this performance penalty?- Is this something other people experienced as well during the initial load of a logical replication with large tables?- We are now thinking of temporarily increasing the number of CPU's and RAM for the migration. Would this help in this case?
I've seen similar symptoms in cases with (a) home-made queues in the tables and (b) long transactions.
Unfortunately, queue requires frequent vacuuming to preserve more or less constant size of the queue and it's indexes.
And long transactions prevent the vacuum from cleaning up the queue. Initial synchronization phase of the logical replication is in fact such a transaction.
I would recommend doing the following:
- avoid adding ALL tables to the publication
- instead, split all tables in a batches in such a way, that initial batch processing takes limited time (say, 15-30 minutes at most)
- of course, this leaves the biggest tables alone — add those one by one to the publication, preferably at the time slot with minimal load on the queue.
- make sure to catch up on the queue processing and vacuum it between batches
- on the receiving side, avoid creating indexes on the tables: create just a necessary PK or UK, wait for the initial load to complete and then add all the rest ones
As for the queue, PGQ from skytools is using different approach to maintain queue tables:
- once in a while (2 hours by default) processing is switched to a new table, tab_1, tab_2, tab_3 are used in a round
- after the switch, any remaining entries can be moved from previous to the live table (shouldn't be necessary if switch is done properly, although might be tricky in a presence of a long transactions)
- previous table is TRUNCATEd
In your case, you can do `VACUUM FULL` between replicating each batch of tables.
Unfortunately, queue requires frequent vacuuming to preserve more or less constant size of the queue and it's indexes.
And long transactions prevent the vacuum from cleaning up the queue. Initial synchronization phase of the logical replication is in fact such a transaction.
I would recommend doing the following:
- avoid adding ALL tables to the publication
- instead, split all tables in a batches in such a way, that initial batch processing takes limited time (say, 15-30 minutes at most)
- of course, this leaves the biggest tables alone — add those one by one to the publication, preferably at the time slot with minimal load on the queue.
- make sure to catch up on the queue processing and vacuum it between batches
- on the receiving side, avoid creating indexes on the tables: create just a necessary PK or UK, wait for the initial load to complete and then add all the rest ones
As for the queue, PGQ from skytools is using different approach to maintain queue tables:
- once in a while (2 hours by default) processing is switched to a new table, tab_1, tab_2, tab_3 are used in a round
- after the switch, any remaining entries can be moved from previous to the live table (shouldn't be necessary if switch is done properly, although might be tricky in a presence of a long transactions)
- previous table is TRUNCATEd
In your case, you can do `VACUUM FULL` between replicating each batch of tables.
Victor Yegorov
Hi,
- on the receiving side, avoid creating indexes on the tables: create just a necessary PK or UK, wait for the initial load to complete and then add all the rest ones
Thanks, this is a good tip. We are going to add this
We also noticed the code that was getting the next from the "queue" was doing the query with 'select for update', but without a 'SKIP LOCKED'. This is probably also something that caused more wait time if the server is more busy as usual during the replication. So we are going to add this and try again. We are also minimizing load on the queue during initial replication.
On to the next try.
Lars
On Wed, Dec 9, 2020 at 6:45 PM Victor Yegorov <vyegorov@gmail.com> wrote:
ср, 9 дек. 2020 г. в 10:21, Lars Vonk <lars.vonk@gmail.com>:--We are doing a logical postgres replication from Postgres 11 to 12. Our database is around 700GB (8 cpu's, 32 GB).During the replication process, at some point, we see a huge performance penalty on a particular table. This table acts as a queue with lots of inserts and deletes happening throughout the day. For most of the time this table is empty, but during this performance penalty the number of rows in this table grows to 10.000 rows, and processing is not fast enough to empty this table. Main reason for this (as far as we see) is that the performance of the query for selecting the next row to process drops from < 10MS to 400MS. This eventually causes too much cpu load on the Primary and we have to cancel the replication process.We already tried the initial load three times, and it consistently fails with the same "error". Last try was a per table approach and excluding this "queue" table.After cancelling the replication the query is fast again and the load on the Primary goes back to normal. We see that this happens when replicating large tables (> millions of rows). During this performance penalty the explain of the query selecting the next row from this table tells us it is doing a sequential scan (there is an index but it is not used).- What could cause this performance penalty?- Is this something other people experienced as well during the initial load of a logical replication with large tables?- We are now thinking of temporarily increasing the number of CPU's and RAM for the migration. Would this help in this case?I've seen similar symptoms in cases with (a) home-made queues in the tables and (b) long transactions.
Unfortunately, queue requires frequent vacuuming to preserve more or less constant size of the queue and it's indexes.
And long transactions prevent the vacuum from cleaning up the queue. Initial synchronization phase of the logical replication is in fact such a transaction.
I would recommend doing the following:
- avoid adding ALL tables to the publication
- instead, split all tables in a batches in such a way, that initial batch processing takes limited time (say, 15-30 minutes at most)
- of course, this leaves the biggest tables alone — add those one by one to the publication, preferably at the time slot with minimal load on the queue.
- make sure to catch up on the queue processing and vacuum it between batches
- on the receiving side, avoid creating indexes on the tables: create just a necessary PK or UK, wait for the initial load to complete and then add all the rest ones
As for the queue, PGQ from skytools is using different approach to maintain queue tables:
- once in a while (2 hours by default) processing is switched to a new table, tab_1, tab_2, tab_3 are used in a round
- after the switch, any remaining entries can be moved from previous to the live table (shouldn't be necessary if switch is done properly, although might be tricky in a presence of a long transactions)
- previous table is TRUNCATEd
In your case, you can do `VACUUM FULL` between replicating each batch of tables.Victor Yegorov
It has been 4 hours and it is safe to say that the measurements we took have a huge positive effect: > 30 times faster and no noticeable effect on the running Primary at all.
A 20GB table is now replicated under 10 minutes.
- We removed all non PK and unique indices from the large tables
- We the changed the query on the queue table to add the 'SKIP LOCKED' clause.
- We do a per table approach for the larger tables.
I think the indices have the most significant impact, but not sure how to proof this since we did multiple changes at the same time.
Thanks again for the tips!
-- Lars
On Thu, Dec 10, 2020 at 9:12 AM Lars Vonk <lars.vonk@gmail.com> wrote:
Hi,- on the receiving side, avoid creating indexes on the tables: create just a necessary PK or UK, wait for the initial load to complete and then add all the rest onesThanks, this is a good tip. We are going to add thisWe also noticed the code that was getting the next from the "queue" was doing the query with 'select for update', but without a 'SKIP LOCKED'. This is probably also something that caused more wait time if the server is more busy as usual during the replication. So we are going to add this and try again. We are also minimizing load on the queue during initial replication.On to the next try.LarsOn Wed, Dec 9, 2020 at 6:45 PM Victor Yegorov <vyegorov@gmail.com> wrote:ср, 9 дек. 2020 г. в 10:21, Lars Vonk <lars.vonk@gmail.com>:--We are doing a logical postgres replication from Postgres 11 to 12. Our database is around 700GB (8 cpu's, 32 GB).During the replication process, at some point, we see a huge performance penalty on a particular table. This table acts as a queue with lots of inserts and deletes happening throughout the day. For most of the time this table is empty, but during this performance penalty the number of rows in this table grows to 10.000 rows, and processing is not fast enough to empty this table. Main reason for this (as far as we see) is that the performance of the query for selecting the next row to process drops from < 10MS to 400MS. This eventually causes too much cpu load on the Primary and we have to cancel the replication process.We already tried the initial load three times, and it consistently fails with the same "error". Last try was a per table approach and excluding this "queue" table.After cancelling the replication the query is fast again and the load on the Primary goes back to normal. We see that this happens when replicating large tables (> millions of rows). During this performance penalty the explain of the query selecting the next row from this table tells us it is doing a sequential scan (there is an index but it is not used).- What could cause this performance penalty?- Is this something other people experienced as well during the initial load of a logical replication with large tables?- We are now thinking of temporarily increasing the number of CPU's and RAM for the migration. Would this help in this case?I've seen similar symptoms in cases with (a) home-made queues in the tables and (b) long transactions.
Unfortunately, queue requires frequent vacuuming to preserve more or less constant size of the queue and it's indexes.
And long transactions prevent the vacuum from cleaning up the queue. Initial synchronization phase of the logical replication is in fact such a transaction.
I would recommend doing the following:
- avoid adding ALL tables to the publication
- instead, split all tables in a batches in such a way, that initial batch processing takes limited time (say, 15-30 minutes at most)
- of course, this leaves the biggest tables alone — add those one by one to the publication, preferably at the time slot with minimal load on the queue.
- make sure to catch up on the queue processing and vacuum it between batches
- on the receiving side, avoid creating indexes on the tables: create just a necessary PK or UK, wait for the initial load to complete and then add all the rest ones
As for the queue, PGQ from skytools is using different approach to maintain queue tables:
- once in a while (2 hours by default) processing is switched to a new table, tab_1, tab_2, tab_3 are used in a round
- after the switch, any remaining entries can be moved from previous to the live table (shouldn't be necessary if switch is done properly, although might be tricky in a presence of a long transactions)
- previous table is TRUNCATEd
In your case, you can do `VACUUM FULL` between replicating each batch of tables.Victor Yegorov