Skip to content

Latest commit

 

History

History
163 lines (117 loc) · 7.24 KB

0052_how_to_reduce_wal_generation_rates.md

File metadata and controls

163 lines (117 loc) · 7.24 KB

Originally from: tweet, LinkedIn post.

How to reduce WAL generation rates

I post a new PostgreSQL "howto" article every day. Join me in this journey – subscribe, provide feedback, share!

In a fast-growing project, one of the important optimization vectors is reduction of the amount of WAL (Write Ahead Log) generated.

Why it is important

WAL is a core mechanism that serves as foundation for post-failure recovery, backups, and replication.

The more WAL is generated per second, the more data needs to be replicated and backed up per second, and, as a consequence, various risks may grow as well: replication lags, WAL archiving lags, and post-failure recovery duration.

How to measure WAL generation rates

When a new transaction creates a new WAL record, it receives LSN (Log Sequence Number). Monitoring of the current LSN position is straightforward:

select pg_current_wal_lsn();

This metric should be present in any Postgres monitoring.

The difference between two LSNs is the number of bytes between two positions, and the math can be done in Postgres if needed – using the pg_lsn data type:

nik=# select pg_size_pretty('3/ED5F1E0'::pg_lsn - '0/110A1E0');
 pg_size_pretty
----------------
 12 GB
(1 row)

If your monitoring doesn't have it, you can understand how much of WAL data was generated per hour or day by looking at:

  • pg_wal directory to see the WAL file names
  • inspecting the backups (for example, checking the names of two full backups created by WAL-G: wal-g backup-list --detail)

Both methods should help you to get two LSN values corresponding to two distant points of time.

For further details, see Day 9: How to understand the LSN values and WAL file name.

WAL metrics in query analysis

Since Postgres 13, both pg_stat_statements and EXPLAIN can provide WAL-related metrics:

  1. pg_stat_statements: metrics wal_records, wal_fpi, wal_bytes (docs). A basic analysis example:

    with time_period(delta_sec) as (
  select extract(epoch from now() - stats_reset)
  from pg_stat_statements_info
)
select
  now(),
  delta_sec,
  round(wal_bytes / delta_sec) as wal_bytes_per_sec,
  round(wal_bytes / calls) as wal_bytes_per_call,
  round(total_exec_time::numeric / delta_sec, 2) as exec_ms_per_sec,
  round(mean_exec_time::numeric, 2) as exec_ms_per_call,
  queryid
from
  pg_stat_statements,
  time_period
order by wal_bytes desc
limit 25;

  2. EXPLAIN: use explain (analyze, buffers, wal) to see the WAL metrics in the execution plan

    nik=# explain (analyze, buffers, wal) insert into t select i from generate_series(1, 100000) as i;
                                                             QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------
 Insert on t  (cost=0.00..1000.00 rows=0 width=0) (actual time=159.378..159.378 rows=0 loops=1)
   Buffers: shared hit=100895 dirtied=442 written=442
   WAL: records=100000 fpi=1 bytes=5900343
   ->  Function Scan on generate_series i  (cost=0.00..1000.00 rows=100000 width=4) (actual time=26.179..30.696 rows=100000 loops=1)
 Planning Time: 1.945 ms
 Execution Time: 160.483 ms
(6 rows)

Full-page writes

The "fpi" metrics in both pg_stat_statements and EXPLAIN results show how many full-page images (full-page writes) happened.

If configuration parameter full_page_write is on (it is so by default; check it: show full_page_writes;), then after each checkpoint the very first change in a page leads to the whole page being written in WAL. By default, the page size is 8 KiB and it is so in most Postgres installations (check it: show block_size;). This means that if only a very small part of the page is changed, still the whole page needs to be written after checkpoint. Subsequent writes in the same page are going to be normal (only changes are recorded to WAL), but once a new checkpoint happens, then again, a new full-page write is needed first. More about it:

Optimization ideas

Below we discuss various ideas that can help you reduce the amount of WAL generated.

  1. Checkpoint tuning: increase distance between checkpoints

    Increasing intervals between checkpoints gives two benefits, especially if workload contains many random (not sequential like during massive data load with COPY) writes:

    • fewer full page writes,
    • fewer repetitive flushes of the same buffer (that once flushed, might become dirty again too quickly due to new writes).

    To increase distance, we just need to increase max_wal_size (default 1GB) and checkpoint_timeout (default 5min). But this needs to be done with understanding of the trade-off: the bigger distance between checkpoints means more WALs will need to be replayed to achieve consistency point in various situations:

    • longer recover time after crashes,
    • longer time to provision new nodes from backups.

    Still, this method is a must-have for larger setups, since it gives substantial improvement.

    Changing max_wal_size and checkpoint_timeout doesn't require restart.

  2. Checkpoint tuning: enable compression of full-page writes

    Consider wal_compression – compression of full-page writes. In most cases, it is worth doing it (although, there are some reports that it led to higher CPU usage and decision to revert the change).

    Changing it doesn't require restart.

  3. Optimize queries

    Using pg_stat_statements and EXPLAIN, as discussed above, find and optimize the most WAL-write-heavy queries.

    One of the approaches to optimize writes is to encourage Postgres to use more HOT UPDATEs (for that, we need to ensure that pages have free space – surprisingly, some bloat is helpful here – and we don't over-index the table so the columns we're changing do not participate in index definitions).

  4. Drop unused and redundant indexes

    During query optimization, keep in mind that for non-HOT UPDATEs and INSERTs, amount of WAL produced depends on the number of indexes the table has. Index cleanup is a very helpful approach to reduce WAL volumes from such writes.

  5. Partitioning

    Partitioning of large (100+ GiB) tables improves data locality for writes – for example, UPDATEs of a bunch of rows could be scattered among many pages if table is not partitioned, and with partitioning schema that defines old partitions (that receive almost no writes) and partitions with fresh data, most writes are going to be localized in the fresh partitions, which can help reduce WAL generation rates.