redis/tests/integration/replication-buffer.tcl
Mincho Paskalev e3c38aab66
Handle primary/replica clients in IO threads (#14335)
# Problem

While introducing Async IO
threads(https://github.com/redis/redis/pull/13695) primary and replica
clients were left to be handled inside main thread due to data race and
synchronization issues. This PR solves this issue with the additional
hope it increases performance of replication.

# Overview

## Moving the clients to IO threads

Since clients first participate in a handshake and an RDB replication
phases it was decided they are moved to IO-thread after RDB replication
is done. For primary client this was trivial as the master client is
created only after RDB sync (+ some additional checks one can see in
`isClientMustHandledByMainThread`). Replica clients though are moved to
IO threads immediately after connection (as are all clients) so
currently in `unstable` replication happens while this client is in
IO-thread. In this PR it was moved to main thread after receiving the
first `REPLCONF` message from the replica, but it is a bit hacky and we
can remove it. I didn't find issues between the two versions.

## Primary client (replica node)

We have few issues here:
- during `serverCron` a `replicationCron` is ran which periodically
sends `REPLCONF ACK` message to the master, also checks for timed-out
master. In order to prevent data races we utilize`IOThreadClientsCron`.
The client is periodically sent to main thread and during
`processClientsFromIOThread` it's checked if it needs to run the
replication cron behaviour.

- data races with main thread - specifically `lastinteraction` and
`read_reploff` members of the primary client that are written to in
`readQueryFromClient` could be accessed at the same time from main
thread during execution of `INFO REPLICATION`(`genRedisInfoString`). To
solve this the members were duplicated so if the client is in IO-thread
it writes to the duplicates and they are synced with the original
variables each time the client is send to main thread ( that means `INFO
REPLICATION` could potentially return stale values).

- During `freeClient` the primary client is fetched to main thread but
when caching it(`replicationCacheMaster`) the thread id will remain the
id of the IO thread it was from. This creates problems when resurrecting
the master client. Here the call to `unbindClientFromIOThreadEventLoop`
in `freeClient` was rewritten to call `keepClientInMainThread` which
automatically fixes the problem.

- During `exitScriptTimedoutMode` the master is queued for reprocessing
(specifically process any pending commands ASAP after it's unblocked).
We do that by putting it in the `server.unblocked_clients` list, which
are processed in the next `beforeSleep` cycle in main thread. Since this
will create a contention between main and IO thread, we just skip this
queueing in `unblocked_clients` and just queue the client to main thread
- the `processClientsFromIOThread` will process the pending commands
just as main would have.

## Replica clients (primary node)

We move the client after RDB replication is done and after replication
backlog is fed with its first message.
We do that so that the client's reference to the first replication
backlog node is initialized before it's read from IO-thread, hence no
contention with main thread on it.

### Shared replication buffer

Currently in unstable the replication buffer is shared amongst clients.
This is done via clients holding references to the nodes inside the
buffer. A node from the buffer can be trimmed once each replica client
has read it and send its contents. The reference is
`client->ref_repl_buf_node`. The replication buffer is written to by
main thread in `feedReplicationBuffer` and the refcounting is intrusive
- it's inside the replication-buffer nodes themselves.

Since the replica client changes the refcount (decreases the refcount of
the node it has just read, and increases the refcount of the next node
it starts to read) during `writeToClient` we have a data race with main
thread when it feeds the replication buffer. Moreover, main thread also
updates the `used` size of the node - how much it has written to it,
compared to its capacity which the replica client relies on to know how
much to read. Obviously replica being in IO-thread creates another data
race here. To mitigate these issues a few new variables were added to
the client's struct:

- `io_curr_repl_node` - starting node this replica is reading from
inside IO-thread
- `io_bound_repl_node` - the last node in the replication buffer the
replica sees before being send to IO-thread.

These values are only allowed to be updated in main thread. The client
keeps track of how much it has read into the buffer via the old
`ref_repl_buf_node`. Generally while in IO-thread the replica client
will now keep refcount of the `io_curr_repl_node` until it's processed
all the nodes up to `io_bound_repl_node` - at that point its returned to
main thread which can safely update the refcounts.
The `io_bound_repl_node` reference is there so the replica knows when to
stop reading from the repl buffer - imagine that replica reads from the
last node of the replication buffer while main thread feeds data to it -
we will create a data race on the `used` value
(`_writeToClientSlave`(IO-thread) vs `feedReplicationBuffer`(main)).
That's why this value is updated just before the replica is being send
to IO thread.
*NOTE*, this means that when replicas are handled by IO threads they
will hold more than one node at a time (i.e `io_curr_repl_node` up to
`io_bound_repl_node`) meaning trimming will happen a bit less
frequently. Tests show no significant problems with that.
(tnx to @ShooterIT for the `io_curr_repl_node` and `io_bound_repl_node`
mechanism as my initial implementation had similar semantics but was way
less clear)

Example of how this works:

* Replication buffer state at time N:
   | node 0| ... | node M, used_size K |
* replica caches `io_curr_repl_node`=0, `io_bound_repl_node`=M and
`io_bound_block_pos`=K
* replica moves to IO thread and processes all the data it sees
* Replication buffer state at time N + 1:
| node 0| ... | node M, used_size Full | |node M + 1| |node M + 2,
used_size L|, where Full > M
* replica moves to main thread at time N + 1, at this point following
happens
   - refcount to node 0 (io_curr_repl_node) is decreased
- `ref_repl_buf_node` becomes node M(io_bound_repl_node) (we still have
size-K bytes to process from there)
- refcount to node M is increased (now all nodes from 0 up to M-1
including can be trimmed unless some other replica holds reference to
them)
- And just before the replica is send back to IO thread the following
are updated:
   - `io_bound_repl_node` ref becomes node M+2
   - `io_bound_block_pos` becomes L

Note that replica client is only moved to main if it has processed all
the data it knows about (i.e up to `io_bound_repl_node` +
`io_bound_block_pos`)

### Replica clients kept in main as much as possible

During implementation an issue arose - how fast is the replica client
able to get knowledge about new data from the replication buffer and how
fast can it trim it. In order for that to happen ASAP whenever a replica
is moved to main it remains there until the replication buffer is fed
new data. At that point its put in the pending write queue and special
cased in handleClientsWithPendingWrites so that its send to IO thread
ASAP to write the new data to replica. Also since each time the replica
writes its whole repl data it knows about that means after it's send to
main thread `processClientsFromIOThread` is able to immediately update
the refcounts and trim whatever it can.

### ACK messages from primary

Slave clients need to periodically read `REPLCONF ACK` messages from
client. Since replica can remain in main thread indefinitely if no DB
change occurs, a new atomic `pending_read` was added during
`readQueryFromClient`. If a replica client has a pending read it's
returned back to IO-thread in order to process the read even if there is
no pending repl data to write.

### Replicas during shutdown

During shutdown the main thread pauses write actions and periodically
checks if all replicas have reached the same replication offset as the
primary node. During `finishShutdown` that may or may not be the case.
Either way a client data may be read from the replicas and even we may
try to write any pending data to them inside `flushSlavesOutputBuffers`.
In order to prevent races all the replicas from IO threads are moved to
main via `fetchClientFromIOThread`. A cancel of the shutdown should be
ok, since the mechanism employed by `handleClientsWithPendingWrites`
should return the client back to IO thread when needed.

## Notes

While adding new tests timing issues with Tsan tests were found and
fixed.

Also there is a data race issue caught by Tsan on the `last_error`
member of the `client` struct. It happens when both IO-thread and main
thread make a syscall using a `client` instance - this can happen only
for primary and replica clients since their data can be accessed by
commands send from other clients. Specific example is the `INFO
REPLICATION` command.
Although other such races were fixed, as described above, this once is
insignificant and it was decided to be ignored in `tsan.sup`.

---------

Co-authored-by: Yuan Wang <wangyuancode@163.com>
Co-authored-by: Yuan Wang <yuan.wang@redis.com>
2026-01-21 16:19:12 +02:00

365 lines
15 KiB
Tcl

#
# Copyright (c) 2009-Present, Redis Ltd.
# All rights reserved.
#
# Copyright (c) 2024-present, Valkey contributors.
# All rights reserved.
#
# Licensed under your choice of (a) the Redis Source Available License 2.0
# (RSALv2); or (b) the Server Side Public License v1 (SSPLv1); or (c) the
# GNU Affero General Public License v3 (AGPLv3).
#
# Portions of this file are available under BSD3 terms; see REDISCONTRIBUTIONS for more information.
#
# This test group aims to test that all replicas share one global replication buffer,
# two replicas don't make replication buffer size double, and when there is no replica,
# replica buffer will shrink.
foreach rdbchannel {"yes" "no"} {
start_server {tags {"repl external:skip"}} {
start_server {} {
start_server {} {
start_server {} {
set replica1 [srv -3 client]
set replica2 [srv -2 client]
set replica3 [srv -1 client]
$replica1 config set repl-rdb-channel $rdbchannel
$replica2 config set repl-rdb-channel $rdbchannel
$replica3 config set repl-rdb-channel $rdbchannel
set master [srv 0 client]
set master_host [srv 0 host]
set master_port [srv 0 port]
$master config set save ""
$master config set repl-backlog-size 16384
$master config set repl-diskless-sync-delay 5
$master config set repl-diskless-sync-max-replicas 1
$master config set client-output-buffer-limit "replica 0 0 0"
$master config set repl-rdb-channel $rdbchannel
# Make sure replica3 is synchronized with master
$replica3 replicaof $master_host $master_port
wait_for_sync $replica3
# Generating RDB will take some 100 seconds
$master config set rdb-key-save-delay 1000000
populate 100 "" 16
# Make sure replica1 and replica2 are waiting bgsave
$master config set repl-diskless-sync-max-replicas 2
$replica1 replicaof $master_host $master_port
$replica2 replicaof $master_host $master_port
wait_for_condition 50 100 {
([s rdb_bgsave_in_progress] == 1) &&
[lindex [$replica1 role] 3] eq {sync} &&
[lindex [$replica2 role] 3] eq {sync}
} else {
fail "fail to sync with replicas"
}
test "All replicas share one global replication buffer rdbchannel=$rdbchannel" {
set before_used [s used_memory]
populate 1024 "" 1024 ; # Write extra 1M data
# In case we are running with IO-threads we need to give a few cycles
# for IO-threads to start sending the cmd stream. If we don't do that
# the checks related to the repl_buf_mem will be incorrect as the buffer
# will still be full with the above 1Mb data.
set iothreads [s io_threads_active]
if {$iothreads && $rdbchannel == "yes"} {
after 1000
}
# New data uses 1M memory, but all replicas use only one
# replication buffer, so all replicas output memory is not
# more than double of replication buffer.
set repl_buf_mem [s mem_total_replication_buffers]
set extra_mem [expr {[s used_memory]-$before_used-1024*1024}]
if {$rdbchannel == "yes"} {
# master's replication buffers should not grow
assert {$extra_mem < 1024*1024}
assert {$repl_buf_mem < 1024*1024}
} else {
assert {$extra_mem < 2*$repl_buf_mem}
}
# Kill replica1, replication_buffer will not become smaller
catch {$replica1 shutdown nosave}
wait_for_condition 50 100 {
[s connected_slaves] eq {2}
} else {
fail "replica doesn't disconnect with master"
}
assert_equal $repl_buf_mem [s mem_total_replication_buffers]
}
test "Replication buffer will become smaller when no replica uses rdbchannel=$rdbchannel" {
# Make sure replica3 catch up with the master
wait_for_ofs_sync $master $replica3
set repl_buf_mem [s mem_total_replication_buffers]
# Kill replica2, replication_buffer will become smaller
catch {$replica2 shutdown nosave}
wait_for_condition 50 100 {
[s connected_slaves] eq {1}
} else {
fail "replica2 doesn't disconnect with master"
}
if {$rdbchannel == "yes"} {
# master's replication buffers should not grow
assert {1024*512 > [s mem_total_replication_buffers]}
} else {
assert {[expr $repl_buf_mem - 1024*1024] > [s mem_total_replication_buffers]}
}
}
}
}
}
}
}
# This test group aims to test replication backlog size can outgrow the backlog
# limit config if there is a slow replica which keep massive replication buffers,
# and replicas could use this replication buffer (beyond backlog config) for
# partial re-synchronization. Of course, replication backlog memory also can
# become smaller when master disconnects with slow replicas since output buffer
# limit is reached.
foreach rdbchannel {"yes" "no"} {
start_server {tags {"repl external:skip debug_defrag:skip"}} {
start_server {} {
start_server {} {
set replica1 [srv -2 client]
set replica1_pid [s -2 process_id]
set replica2 [srv -1 client]
set replica2_pid [s -1 process_id]
set master [srv 0 client]
set master_host [srv 0 host]
set master_port [srv 0 port]
$master config set save ""
$master config set repl-backlog-size 16384
$master config set repl-rdb-channel $rdbchannel
$master config set client-output-buffer-limit "replica 0 0 0"
# Executing 'debug digest' on master which has many keys costs much time
# (especially in valgrind), this causes that replica1 and replica2 disconnect
# with master.
$master config set repl-timeout 1000
$replica1 config set repl-timeout 1000
$replica1 config set repl-rdb-channel $rdbchannel
$replica1 config set client-output-buffer-limit "replica 1024 0 0"
$replica2 config set repl-timeout 1000
$replica2 config set client-output-buffer-limit "replica 1024 0 0"
$replica2 config set repl-rdb-channel $rdbchannel
$replica1 replicaof $master_host $master_port
wait_for_sync $replica1
test "Replication backlog size can outgrow the backlog limit config rdbchannel=$rdbchannel" {
# Generating RDB will take 1000 seconds
$master config set rdb-key-save-delay 1000000
populate 1000 master 10000
$replica2 replicaof $master_host $master_port
# Make sure replica2 is waiting bgsave
wait_for_condition 5000 100 {
([s rdb_bgsave_in_progress] == 1) &&
[lindex [$replica2 role] 3] eq {sync}
} else {
fail "fail to sync with replicas"
}
# Replication actual backlog grow more than backlog setting since
# the slow replica2 kept replication buffer.
populate 20000 master 10000
assert {[s repl_backlog_histlen] > [expr 10000*10000]}
}
# Wait replica1 catch up with the master
wait_for_condition 1000 100 {
[s -2 master_repl_offset] eq [s master_repl_offset]
} else {
fail "Replica offset didn't catch up with the master after too long time"
}
test "Replica could use replication buffer (beyond backlog config) for partial resynchronization rdbchannel=$rdbchannel" {
# replica1 disconnects with master
$replica1 replicaof [srv -1 host] [srv -1 port]
# Write a mass of data that exceeds repl-backlog-size
populate 10000 master 10000
# replica1 reconnects with master
$replica1 replicaof $master_host $master_port
wait_for_condition 1000 100 {
[s -2 master_repl_offset] eq [s master_repl_offset]
} else {
fail "Replica offset didn't catch up with the master after too long time"
}
# replica2 still waits for bgsave ending
assert {[s rdb_bgsave_in_progress] eq {1} && [lindex [$replica2 role] 3] eq {sync}}
# master accepted replica1 partial resync
assert_equal [s sync_partial_ok] {1}
assert_equal [$master debug digest] [$replica1 debug digest]
}
test "Replication backlog memory will become smaller if disconnecting with replica rdbchannel=$rdbchannel" {
assert {[s repl_backlog_histlen] > [expr 2*10000*10000]}
assert_equal [s connected_slaves] {2}
pause_process $replica2_pid
r config set client-output-buffer-limit "replica 128k 0 0"
# trigger output buffer limit check
r set key [string repeat A [expr 64*1024]]
# master will close replica2's connection since replica2's output
# buffer limit is reached, so there only is replica1.
# In case of rdbchannel=yes, main channel will be disconnected only.
wait_for_condition 100 100 {
[s connected_slaves] eq {1} ||
([s connected_slaves] eq {2} &&
[string match {*slave*state=wait_bgsave*} [$master info]])
} else {
fail "master didn't disconnect with replica2"
}
# Since we trim replication backlog inrementally, replication backlog
# memory may take time to be reclaimed.
wait_for_condition 1000 100 {
[s repl_backlog_histlen] < [expr 10000*10000]
} else {
fail "Replication backlog memory is not smaller"
}
resume_process $replica2_pid
}
# speed up termination
$master config set shutdown-timeout 0
}
}
}
}
foreach rdbchannel {"yes" "no"} {
test "Partial resynchronization is successful even client-output-buffer-limit is less than repl-backlog-size rdbchannel=$rdbchannel" {
start_server {tags {"repl external:skip"}} {
start_server {} {
r config set save ""
r config set repl-backlog-size 100mb
r config set client-output-buffer-limit "replica 512k 0 0"
r config set repl-rdb-channel $rdbchannel
set replica [srv -1 client]
$replica config set repl-rdb-channel $rdbchannel
$replica replicaof [srv 0 host] [srv 0 port]
wait_for_sync $replica
set big_str [string repeat A [expr 10*1024*1024]] ;# 10mb big string
r multi
r client kill type replica
r set key $big_str
r set key $big_str
r debug sleep 2 ;# wait for replica reconnecting
r exec
# When replica reconnects with master, master accepts partial resync,
# and don't close replica client even client output buffer limit is
# reached.
r set key $big_str ;# trigger output buffer limit check
wait_for_ofs_sync r $replica
# master accepted replica partial resync
assert_equal [s sync_full] {1}
assert_equal [s sync_partial_ok] {1}
r multi
r set key $big_str
r set key $big_str
r exec
# replica's reply buffer size is more than client-output-buffer-limit but
# doesn't exceed repl-backlog-size, we don't close replica client.
wait_for_condition 1000 100 {
[s -1 master_repl_offset] eq [s master_repl_offset]
} else {
fail "Replica offset didn't catch up with the master after too long time"
}
assert_equal [s sync_full] {1}
assert_equal [s sync_partial_ok] {1}
}
}
}
# This test was added to make sure big keys added to the backlog do not trigger psync loop.
test "Replica client-output-buffer size is limited to backlog_limit/16 when no replication data is pending rdbchannel=$rdbchannel" {
proc client_field {r type f} {
set client [$r client list type $type]
if {![regexp $f=(\[a-zA-Z0-9-\]+) $client - res]} {
error "field $f not found for in $client"
}
return $res
}
start_server {tags {"repl external:skip"}} {
start_server {} {
set replica [srv -1 client]
set replica_host [srv -1 host]
set replica_port [srv -1 port]
set master [srv 0 client]
set master_host [srv 0 host]
set master_port [srv 0 port]
$master config set maxmemory-policy allkeys-lru
$master config set repl-backlog-size 16384
$master config set client-output-buffer-limit "replica 32768 32768 60"
$master config set repl-rdb-channel $rdbchannel
$replica config set repl-rdb-channel $rdbchannel
# Key has has to be larger than replica client-output-buffer limit.
set keysize [expr 256*1024]
$replica replicaof $master_host $master_port
wait_for_condition 50 100 {
[lindex [$replica role] 0] eq {slave} &&
[string match {*master_link_status:up*} [$replica info replication]]
} else {
fail "Can't turn the instance into a replica"
}
# Write a big key that is gonna breach the obuf limit and cause the replica to disconnect,
# then in the same event loop, add at least 16 more keys, and enable eviction, so that the
# eviction code has a chance to call flushSlavesOutputBuffers, and then run PING to trigger the eviction code
set _v [prepare_value $keysize]
$master write "[format_command mset key $_v k1 1 k2 2 k3 3 k4 4 k5 5 k6 6 k7 7 k8 8 k9 9 ka a kb b kc c kd d ke e kf f kg g kh h]config set maxmemory 1\r\nping\r\n"
$master flush
$master read
$master read
$master read
wait_for_ofs_sync $master $replica
# Write another key to force the test to wait for another event loop iteration so that we
# give the serverCron a chance to disconnect replicas with COB size exceeding the limits
$master config set maxmemory 0
$master set key1 1
wait_for_ofs_sync $master $replica
assert {[status $master connected_slaves] == 1}
wait_for_condition 50 100 {
[client_field $master replica tot-mem] < $keysize
} else {
fail "replica client-output-buffer usage is higher than expected."
}
# now we expect the replica to re-connect but fail partial sync (it doesn't have large
# enough COB limit and must result in a full-sync)
assert {[status $master sync_partial_ok] == 0}
# Before this fix (#11905), the test would trigger an assertion in 'o->used >= c->ref_block_pos'
test {The update of replBufBlock's repl_offset is ok - Regression test for #11666} {
set rd [redis_deferring_client]
set replid [status $master master_replid]
set offset [status $master repl_backlog_first_byte_offset]
$rd psync $replid $offset
assert_equal {PONG} [$master ping] ;# Make sure the master doesn't crash.
$rd close
}
}
}
}
}