AUTOVACUUM to prevent wraparound

autovacuum_freeze_max_age 是 AUTOVACUUM 最不常用的參數，也基本不需要優化，但卻是 AUTOVACUUM 最重要的一個參數，因為它與 wraparound 有關，即使 AUTOVACUUM 關閉，達到這個閾值，也會強制觸發 AUTOVACUUM ，可見它有多重要。

#autovacuum_freeze_max_age = 200000000  # maximum XID age before forced vacuum
                                        # (change requires restart)

首先打開 AUTOVACUUM。為了方便測試，將 autovacuum_freeze_max_age 設置為其最小值。

autovacuum_freeze_max_age = 100000

正常情況是達不到 autovacuum_freeze_max_age 閾值的，因為還沒有達到這個閾值，就觸發了 AUTOVACUUM。為了能測試到這個閾值，人工創建一條 idle in transaction，以阻塞正常的 AUTOVACUUM。

alvindb=# BEGIN;
BEGIN
alvindb=# SELECT txid_current();
 txid_current
--------------
     15200695
(1 row)

alvindb=# UPDATE alvin.tb_test_vacuum SET test_num = test_num WHERE test_num = 1;
UPDATE 1

查到 idle in transaction 如下，可以看到其 backend_xid 為 15200695：

postgres=# SELECT * FROM pg_stat_activity WHERE state = 'idle in transaction';
-[ RECORD 1 ]----+------------------------------
datid            | 37509
datname          | alvindb
pid              | 28497
usesysid         | 10
usename          | postgres
application_name | psql
client_addr      | 
client_hostname  | 
client_port      | -1
backend_start    | 2021-11-13 23:18:33.480814+08
xact_start       | 2021-11-13 23:19:45.288275+08
query_start      | 2021-11-13 23:19:58.133248+08
state_change     | 2021-11-13 23:19:58.133412+08
wait_event_type  | Client
wait_event       | ClientRead
state            | idle in transaction
backend_xid      | 15200695
backend_xmin     | 
query            | SELECT txid_current();
backend_type     | client backend

不斷更新表，使其達到閾值。

alvindb=# SELECT c.relname, c.relfrozenxid,age(c.relfrozenxid) FROM pg_class c WHERE relname = 'pgbench_accounts';
     relname      | relfrozenxid |  age   
------------------+--------------+--------
 pgbench_accounts |     15250696 | 100000
(1 row)

此時修改最后一條數據，使表的 age 超過閾值。

alvindb=# SELECT c.relname, age(c.relfrozenxid) FROM pg_class c WHERE relname = 'pgbench_accounts';
     relname      |  age   
------------------+--------
 pgbench_accounts | 100001
(1 row)

可以看到，此時已觸發 aggressive vacuum to prevent wraparound。

其 query 為 VACUUM public.pgbench_accounts (to prevent wraparound)，並且 backend_xmin 為 15200695。

-[ RECORD 1 ]----+-------------------------------------------------------------------
datid            | 37509
datname          | alvindb
pid              | 30288
usesysid         | 
usename          | 
application_name | 
client_addr      | 
client_hostname  | 
client_port      | 
backend_start    | 2021-11-13 23:30:39.767723+08
xact_start       | 2021-11-13 23:30:39.788839+08
query_start      | 2021-11-13 23:30:39.788839+08
state_change     | 2021-11-13 23:30:39.788839+08
wait_event_type  | 
wait_event       | 
state            | active
backend_xid      | 
backend_xmin     | 15200695
query            | autovacuum: VACUUM public.pgbench_accounts (to prevent wraparound)
backend_type     | autovacuum worker

從 PostgreSQL 日志中可以看到觸發的 automatic aggressive vacuum。

[    2021-11-13 23:30:40323 CST 30288 618fda1f.7650 3 6/72694 0]LOG:  automatic aggressive vacuum of table "alvindb.public.pgbench_accounts": index scans: 1
        pages: 0 removed, 404 remain, 0 skipped due to pins, 0 skipped frozen
        tuples: 3 removed, 100000 remain, 0 are dead but not yet removable, oldest xmin: 15200695
        buffer usage: 438 hits, 480 misses, 412 dirtied
        avg read rate: 28.065 MB/s, avg write rate: 24.089 MB/s
        system usage: CPU: user: 0.02 s, system: 0.03 s, elapsed: 0.53 s
[    2021-11-13 23:30:40.334 CST 30288 618fda1f.7650 4 6/72696 0]WARNING:  oldest xmin is far in the past

所以，如果一旦發現 aggressive vacuum (to prevent wraparound)，需要格外重視，檢查是否出現類似 idle in transaction 等異常情況。

尤其是 PostgreSQL 9.6 之前，還沒有參數 idle_in_transaction_session_timeout(since PostgreSQL 9.6) 的版本，需要手動 terminate idle in transaction。或 PostgreSQL 9.6 之后，但未設置該參數。