Implementing failover

of logical replication
slots in Patroni.
PGConf.DE 2022, Leipzig
Alexander Kukushkin
2022-05-13
About me

Alexander Kukushkin
Database Engineer @ZalandoTech

The Patroni guy

[email protected]

Twitter: @cyberdemn

2
Streaming replication

● Physical
○ replicates binary changes via WAL files over TCP to
standby
○ can’t replicate between different major versions or
platforms
● Logical
○ replicates only data objects and their changes
○ works between different major versions and platforms
○ requires the replication slot
3
Enabling logical replication

● wal_level = logical # default value is replica

● max_wal_senders = 10
● max_replication_slots = 10

Postgres restart is required!

4
Replication slots

● Provide guarantees that WAL segments are not removed

until consumed
● Provide protection against relevant rows being removed by
(auto)vacuum
● Logical replication slots are tightly coupled with particular
database.

5
Creating a logical replication slot

● SELECT pg_create_logical_replication_slot(
'<slot_name>', '<output_plugin_name>');

–- or —-

● CREATE_REPLICATION_SLOT <slot_name>
LOGICAL <output_plugin_name>
○ pg_recvlogical --create-slot --slot=<slot_name> \
--plugin=<plugin_name> --dbname=<database_name>

6
Logical Decoding Plugins

● Built-in core
○ test_decoding
○ pgoutput (logical replication)
● 3rd party
○ wal2json - JSON
○ pglogical
○ Debezium decoderbufs - Protobuf
○ decoder_raw - as SQL
○ …
7
What is Logical Decoding?

● Extracting data changes in a “simple” format that could be

interpreted by an external tool.

● Outside of PostgreSQL it is also known as Change Data

Capture (CDC)

8
Example: creating the slot

localhost/testdb=# SELECT * FROM pg_create_logical_replication_slot

('my_slot', 'test_decoding');

slot_name │ lsn
───────────┼────────────
my_slot │ 0/130694C8
(1 row)

localhost/testdb=# SELECT slot_name, plugin, slot_type, database, confirmed_flush_lsn

FROM pg_replication_slots;

slot_name │ plugin │ slot_type │ database │ confirmed_flush_lsn

───────────┼───────────────┼───────────┼──────────┼─────────────────────
my_slot │ test_decoding │ logical │ testdb │ 0/130694C8
(1 row)
9
Example: peek changes without consuming

localhost/testdb=# CREATE TABLE replicate_me (

id BIGINT NOT NULL GENERATED ALWAYS AS IDENTITY PRIMARY KEY,
name TEXT);
CREATE TABLE

localhost/testdb=# INSERT INTO replicate_me (name) VALUES ('PGConf.DE ');

INSERT 0 1

localhost/testdb=# SELECT * FROM pg_logical_slot_peek_changes( 'my_slot' , NULL, NULL);

lsn │ xid │ data
───────────┼─────┼───────────────────────────────────────────────────────────────────────
0/130694F8 │ 830 │ BEGIN 830
0/1309A768 │ 830 │ COMMIT 830
0/1309A7A0 │ 831 │ BEGIN 831
0/1309A808 │ 831 │ table public.replicate_me: INSERT: id[bigint]:1 name[text]:'PGConf.DE'
0/1309A960 │ 831 │ COMMIT 831

(5 rows)

10
Example: pg_recvlogical

$ pg_recvlogical -h localhost -U postgres -d testdb --slot=my_slot --start -f

-

BEGIN 830

COMMIT 830

BEGIN 831

table public.replicate_me: INSERT: id[bigint]:1 name[text]:'PGConf.DE'

COMMIT 831

^Cpg_recvlogical: error: unexpected termination of replication stream:

11
Logical replication example

CREATE PUBLICATION my_publication

FOR TABLE replicate_me WITH (publish = 'insert');

CREATE SUBSCRIPTION my_sub

CONNECTION 'host=172.168.18.50 port=5432 user=repl dbname=testdb'

PUBLICATION my_publication;

12
Logical replication slots and HA

● Replication slots are not replicated!

● Logical replication slots can’t be created on standbys:

localhost/testdb=# SELECT * FROM

pg_create_logical_replication_slot ('my_slot', 'test_decoding');

ERROR: logical decoding cannot be used while in recovery

● Consumers can’t continue work after failover/switchover.

13
Naive solution in Patroni

● Don’t allow incoming connections after failover/switchover

before logical slots are created:
○ Patroni REST API health-check returns 503
○ Delay callbacks
○ K8s leader Service without endpoints

● Logical events could be silently lost if consumer wasn’t

running or was lagging!
14
Can we create logical slots in the past?

● Yes, with the custom extension: https://github.com/x4m/pg_tm_aux

● Pros: events won't be lost
● Cons: It’s not always possible to install 3rd party extensions
● Potential problems:
○ WAL isn’t accessible
○ May fail to take a “catalog snapshot”

15
 Can we do better?

16
Observation

If we restart Postgres in read-only mode, existing replication slots are still there:
localhost/testdb=# SELECT pg_is_in_recovery();
pg_is_in_recovery
───────────────────
t
(1 row)

localhost/testdb=# SELECT slot_name, plugin, slot_type, database, confirmed_flush_lsn

FROM pg_replication_slots;
slot_name │ plugin │ slot_type │ database │ confirmed_flush_lsn
──────────┼───────────────┼───────────┼──────────┼─────────────────────
my_slot │ test_decoding │ logical │ testdb │ 0/1309AA80
(1 row)

17
 Logical replication slots
can exist on replicas!

18
How replication slots are stored?

$ ls -l $PGDATA/pg_replslot/my_slot/
total 4
-rw------- 1 postgres postgres 200 Apr 11 08:15 state

19
What if we copy the slot file?

● Stop postgres on the replica

● Copy the slot file from the primary

● Start postgres on the replica

● Profit?!
20
Adding the secret sauce

● pg_replication_slot_advance ( slot_name name, upto_lsn pg_lsn )

○ Introduced in PostgreSQL v11 (released in October 2018)

○ Works with logical slots! (must be connected to the right DB)

○ Also works on replicas!

21
Example: pg_replication_slot_advance()

localhost/testdb=# SELECT pg_is_in_recovery();

pg_is_in_recovery
──────────────────
t
(1 row)

localhost/testdb=# SELECT * FROM pg_replication_slot_advance('my_slot', '0/1412FA78');

slot_name │ end_lsn
──────────┼────────────
my_slot │ 0/1412FA78
(1 row)

22
Is it safe?

● Copy the slot file - Yes

● pg_replication_slot_advance() - Yes
● Use replication slot after promote - might be unsafe
localhost/testdb=# SELECT slot_name, plugin, slot_type, database,
catalog_xmin, restart_lsn, confirmed_flush_lsn FROM pg_replication_slots;

slot_name │ plugin │ slot_type │ database │ catalog_xmin │ restart_lsn │ confirmed_flush_lsn

──────────┼───────────────┼───────────┼──────────┼───────────────┼─────────────┼───────────────────
my_slot │ test_decoding │ logical │ testdb │ 831 │ 0/1309A768 │ 0/1309AA80

(1 row)

23
catalog_xmin

● Logical decoding uses pg_catalog to figure out table

structures
○ Needs access to old snapshots
● Race conditions:
○ Slots on replicas are usually behind
○ Autovacuum might clean up pg_catalog tuples required
for decoding
24
How to protect from it?

1. Replicas must use replication slots

(primary_slot_name)

2. hot_standby_feedback must be set to on

25
hot_standby_feedback = off

localhost/ testdb=# SELECT slot_name, plugin, slot_type, database,

catalog_xmin, restart_lsn, confirmed_flush_lsn FROM pg_replication_slots;

Primary:
slot_name │ plugin │ slot_type │ database │ catalog_xmin │ restart_lsn │ confirmed_flush_lsn
────────────┼───────────────┼───────────┼──────────┼──────────────┼─────────────┼────────────────────
my_slot │ test_decoding │ logical │ testdb │ 864 │ 0/14116228 │ 0/1412E410
postgresql1 │ │ physical │ │ │ 0/1412E4C0 │
(2 rows)

Replica:
slot_name │ plugin │ slot_type │ database │ catalog_xmin │ restart_lsn │ confirmed_flush_lsn
──────────┼───────────────┼───────────┼──────────┼──────────────┼─────────────┼─────────────────────
my_slot │ test_decoding │ logical │ testdb │ 863 │ 0/140FE578 │ 0/140FE5B0
(1 row)

26
hot_standby_feedback = on

localhost/ testdb=# SELECT slot_name, plugin, slot_type, database,

catalog_xmin, restart_lsn, confirmed_flush_lsn FROM pg_replication_slots;

Primary:
slot_name │ plugin │ slot_type │ database │ catalog_xmin │ restart_lsn │ confirmed_flush_lsn
────────────┼───────────────┼───────────┼──────────┼──────────────┼─────────────┼────────────────────
my_slot │ test_decoding │ logical │ testdb │ 864 │ 0/14116228 │ 0/1412E410
postgresql1 │ │ physical │ │ 863 │ 0/1412E4C0 │
(2 rows)

Replica:
slot_name │ plugin │ slot_type │ database │ catalog_xmin │ restart_lsn │ confirmed_flush_lsn
──────────┼───────────────┼───────────┼──────────┼──────────────┼─────────────┼─────────────────────
my_slot │ test_decoding │ logical │ testdb │ 863 │ 0/140FE578 │ 0/140FE5B0
(1 row)

27
The plan

1. Copy logical slots to replicas with Postgres restart

a. Fsync files and directories after coping!

2. Periodically call pg_replication_slot_advance()

3. Handle possible errors?

28
Implementation details in Patroni

● Leader: periodically publish confirmed_flush_lsn for all

permanents slots to the Distributed Configuration Store (DCS)
● Replicas:
○ Use pg_read_binary_file() function to copy the slot file if it is
missing on the replica
■ Requires superuser or specially configured rewind_user
○ Enable hot_standby_feedback
29
Cascading replication

The node2 doesn’t have logical slots, but must have

hot_standby_feedback enabled (automatically configured by Patroni).

node1 node2 node3

Primary (nofailover) (replicatefrom:
node2)

30
Possible problems

● Requested WAL segment pg_wal/XXX has already been

removed

● The logical slot could be unsafe to use after promote if

replicas physical slot didn’t reach the catalog_xmin of the
logical slot on the old primary (Patroni shows a warning)

31
Configuring Patroni

$ patronictl edit-config
---
+++
@@ -1,6 +1,12 @@
loop_wait: 10
maximum_lag_on_failover: 1048576
postgresql:
+ use_slots: true
+ parameters:
+ wal_level: logical
use_pg_rewind: true
retry_timeout: 10
ttl: 30
+permanent_slots:
+ my_slot:
+ database: testdb
+ plugin: test_decoding
Apply these changes? [y/N]: y
Configuration changed
32
Logical client issues

● Clients must be prepared to receive some events for the second time after
failover/switchover
● Logical replication may be ahead of physical replication that is acknowledged by
synchronous replicas and see some events that didn’t survive failover.
● Debezium doesn’t correctly handle keepalive messages: DBZ-4055
○ May cause slow (or indefinite) shutdown of the primary because it keeps
walsender process alive
○ “Breaks” switchover in the old (before 2.1.2) Patroni

33
Monitoring

localhost/postgres=# SELECT slot_name,

pg_size_pretty(CASE WHEN pg_is_in_recovery() THEN pg_last_wal_replay_lsn()
ELSE pg_current_wal_lsn() END - confirmed_flush_lsn)
FROM pg_replication_slots WHERE slot_type = 'logical';

slot_name │ pg_size_pretty
──────────┼────────────────
my_slot │ 5736 bytes
(1 row)

34
Conclusion

● Failover of logical slots is supported by Patroni starting

from 2.1.0 (released in July 2021)
○ Requires PostgreSQL v11 or newer
○ The old “feature” (create logical slots after promote) is
disabled.
● Used in production at Zalando for more than 30 databases
● Don’t forget about the monitoring.
35
Credits

● Craig Ringer

36
 Thank you!
Questions?

37