Hi Ramakrishna, 4GB of WAL generated per minute is a lot. I would expect the replay on the subscriber to lag behind because it is a single process. PostgreSQL 16 can create parallel workers for large transactions, however if there is a flood of small transactions touching many tables the single LR worker is going to fall behind.
The only option is split the LR into multiple Publications and Subscriptions as a single worker can't keep up. What is the justification to not split the tables across multiple Publications and Subscriptions Additional items to check Make sure the Subscriber is using binary mode, this avoids an encoding step. https://www.postgresql.org/docs/current/sql-createsubscription.html Avoid the use of IDENTITY SET TO FULL on the publisher, if you do use IDENTITY FULL make sure the subscriber table identity is set to a qualifying unique index. In previous versions of PG the publisher and subscriber identities had to match... IDENTITY SET TO FULL increase the size of the WAL and the work the publisher and subscriber has to do. Hope this helps. On Sat, Sep 21, 2024 at 3:08 PM Ramakrishna m <ram.p...@gmail.com> wrote: > Hi Team, > > We have configured bidirectional replication (but traffic can only flow in > one direction) between two data centers (distance: 1000 km, maximum Network > latency: 100 ms) with an application TPS (transactions per second) of 700 > at maximum. > > We are fine with handling up to 500 TPS without observing any lag between > the two data centers. However, when TPS increases, we notice a lag in WAL > files of over 100 GB (initially, it was 1 TB, but after tuning, it was > reduced to 100 GB). During peak times, WAL files are generated at a rate of > 4 GB per minute. > > All transactions (Tx) take less than 200 ms, with a maximum of 1 second at > times (no long-running transactions). > > *Here are the configured parameters and resources:* > > - *OS*: Ubuntu > - *RAM*: 376 GB > - *CPU*: 64 cores > - *Swap*: 32 GB > - *PostgreSQL Version*: 16.4 (each side has 3 nodes with Patroni and > etcd configured) > - *DB Size*: 15 TB > > *Parameters configured on both sides:* > Name > Setting Unit > > > log_replication_commands off > logical_decoding_work_mem 524288 kB > max_logical_replication_workers 16 > max_parallel_apply_workers_per_subscription 2 > max_replication_slots 20 > max_sync_workers_per_subscription 2 > max_wal_senders 20 > max_worker_processes 40 > wal_level logical > wal_receiver_timeout 600000 ms > wal_segment_size 1073741824 B > wal_sender_timeout 600000 ms > > *Optimizations applied:* > > 1. Vacuum freeze is managed during off-hours; no aggressive vacuum is > triggered during business hours. > 2. Converted a few tables to unlogged. > 3. Removed unwanted tables from publication. > 4. Partitioned all large tables. > > *Pending:* > > 1. Turning off/tuning autovacuum parameters to avoid triggering during > business hours. > > *Not possible: *We are running all tables in a single publication, and it > is not possible to separate them. > > I would greatly appreciate any suggestions you may have to help avoid > logical replication delays, whether through tuning database or operating > system parameters, or any other recommendations > > -- > Thanks & Regards, > Ram. >
