Autovacuum 队列和进度

在许多工作负载下,需要调整 autovacuum 配置,才能跟上死元组的积累速度,默认配置尤其如此。判断当前配置是否足够,可以先比较 autovacuum_max_workers 与实际运行的工作进程数:

SQL
show autovacuum_max_workers;

select
  state,
  count(*),
  array_agg(left(query, 25) order by xact_start)
from pg_stat_activity
where backend_type = 'autovacuum worker'
group by state;

如果活跃工作进程数经常达到 autovacuum_max_workers,就应该考虑提高并发上限,或通过调整成本限制与延迟([auto]vacuum_vacuum_cost_limit / [auto]vacuum_vacuum_cost_delay)提高处理速度。修改参数是否需要重启,应以所用版本中 pg_settings.context 的值为准。

还可以进一步观察:当前有多少张表在“队列”中等待 autovacuum?待处理表的数量有助于判断积压规模,以及现有配置能否跟上。将这一数量与工作进程数对照,可以得到类似 CPU load average 的粗略参考。

以下是回答这个问题的报告 (来源),包括:

它将这些信息进行比较,并构建需要清理的表的列表。

此报告衍生自此处。

另外,它还检查 pg_stat_progress_vacuum 以分析当前正在处理的表。

后续该查询可进一步扩展,以分析"需要自动分析"的表。

SQL
with table_opts as (
  select
    pg_class.oid,
    relname,
    nspname,
    array_to_string(reloptions, '') as relopts
  from pg_class
  join pg_namespace ns on relnamespace = ns.oid
), vacuum_settings as (
  select
    oid,
    relname,
    nspname,
    case
      when relopts like '%autovacuum_vacuum_threshold%' then
        regexp_replace(relopts, '.*autovacuum_vacuum_threshold=([0-9.]+).*', e'\\1')::int8
      else current_setting('autovacuum_vacuum_threshold')::int8
    end as autovacuum_vacuum_threshold,
    case
      when relopts like '%autovacuum_vacuum_scale_factor%'
        then regexp_replace(relopts, '.*autovacuum_vacuum_scale_factor=([0-9.]+).*', e'\\1')::numeric
      else current_setting('autovacuum_vacuum_scale_factor')::numeric
    end as autovacuum_vacuum_scale_factor,
    case
      when relopts ~ 'autovacuum_enabled=(false|off)' then false
      else true
    end as autovacuum_enabled
  from table_opts
), p as (
  select *
  from pg_stat_progress_vacuum
)
select
  coalesce(
    coalesce(nullif(vacuum_settings.nspname, 'public') || '.', '') || vacuum_settings.relname, -- current DB
    format('[something in "%I"]', p.datname) -- another DB
  ) as relation,
  round((100 * psat.n_dead_tup::numeric / nullif(pg_class.reltuples, 0))::numeric, 2) as dead_tup_pct,
  pg_class.reltuples::numeric,
  psat.n_dead_tup,
  format (
    'vt: %s, vsf: %s, %s', -- 'vt' – vacuum_threshold, 'vsf' - vacuum_scale_factor
    vacuum_settings.autovacuum_vacuum_threshold,
    vacuum_settings.autovacuum_vacuum_scale_factor,
    (case when autovacuum_enabled then 'enabled' else 'DISABLED' end)
  ) as effective_settings,
  case
    when last_autovacuum > coalesce(last_vacuum, '0001-01-01') then left(last_autovacuum::text, 19) || ' (auto)'
    when last_vacuum is not null then left(last_vacuum::text, 19) || ' (manual)'
    else null
  end as last_vacuumed,
  coalesce(p.phase, '~~~ in queue ~~~') as status,
  p.pid as pid,
  case
    when a.query ~ '^autovacuum.*to prevent wraparound' then 'wraparound' 
    when a.query ~ '^vacuum' then 'user'
    when a.pid is null then null
    else 'regular'
  end as mode,
  case
    when a.pid is null then null
    else coalesce(wait_event_type || '.' || wait_event, 'f')
  end as waiting,
  round(100.0 * p.heap_blks_scanned / nullif(p.heap_blks_total, 0), 1) as scanned_pct,
  round(100.0 * p.heap_blks_vacuumed / nullif(p.heap_blks_total, 0), 1) as vacuumed_pct,
  p.index_vacuum_count,
  case 
    when psat.relid is not null and p.relid is not null then
      (select count(*) from pg_index where indrelid = psat.relid)
    else null
  end as index_count
from pg_stat_all_tables psat
join pg_class on psat.relid = pg_class.oid
left join vacuum_settings on pg_class.oid = vacuum_settings.oid
full outer join p on p.relid = psat.relid and p.datname = current_database()
left join pg_stat_activity a using (pid)
where
  psat.relid is null
  or p.phase is not null
  or (
    autovacuum_vacuum_threshold
      + (autovacuum_vacuum_scale_factor::numeric * pg_class.reltuples)
    < psat.n_dead_tup
  )
order by status, relation;

这里的“队列”是根据阈值和统计信息推算出的待清理表集合,并非 autovacuum 调度器内部队列的精确快照。上游也建议后续扩展该查询,纳入需要 auto-analyze 的表。

输出示例(在 psql 中运行,并用 \gx 代替末尾的 ;):

tables to be autovacuumed

讨论与补充

在 GitHub 查看讨论

支持作者

感谢支持 PostgreSQL Howto 的长期维护。

微信赞赏码

读者交流群

二维码若已过期,可通过公众号 PostgreSQL 学徒或邮件联系。

PostgreSQL 学徒读者交流群二维码