### Description
A mechanism for disconnecting clients when the sum of all connected clients is above a
configured limit. This prevents eviction or OOM caused by accumulated used memory
between all clients. It's a complimentary mechanism to the `client-output-buffer-limit`
mechanism which takes into account not only a single client and not only output buffers
but rather all memory used by all clients.
#### Design
The general design is as following:
* We track memory usage of each client, taking into account all memory used by the
client (query buffer, output buffer, parsed arguments, etc...). This is kept up to date
after reading from the socket, after processing commands and after writing to the socket.
* Based on the used memory we sort all clients into buckets. Each bucket contains all
clients using up up to x2 memory of the clients in the bucket below it. For example up
to 1m clients, up to 2m clients, up to 4m clients, ...
* Before processing a command and before sleep we check if we're over the configured
limit. If we are we start disconnecting clients from larger buckets downwards until we're
under the limit.
#### Config
`maxmemory-clients` max memory all clients are allowed to consume, above this threshold
we disconnect clients.
This config can either be set to 0 (meaning no limit), a size in bytes (possibly with MB/GB
suffix), or as a percentage of `maxmemory` by using the `%` suffix (e.g. setting it to `10%`
would mean 10% of `maxmemory`).
#### Important code changes
* During the development I encountered yet more situations where our io-threads access
global vars. And needed to fix them. I also had to handle keeps the clients sorted into the
memory buckets (which are global) while their memory usage changes in the io-thread.
To achieve this I decided to simplify how we check if we're in an io-thread and make it
much more explicit. I removed the `CLIENT_PENDING_READ` flag used for checking
if the client is in an io-thread (it wasn't used for anything else) and just used the global
`io_threads_op` variable the same way to check during writes.
* I optimized the cleanup of the client from the `clients_pending_read` list on client freeing.
We now store a pointer in the `client` struct to this list so we don't need to search in it
(`pending_read_list_node`).
* Added `evicted_clients` stat to `INFO` command.
* Added `CLIENT NO-EVICT ON|OFF` sub command to exclude a specific client from the
client eviction mechanism. Added corrosponding 'e' flag in the client info string.
* Added `multi-mem` field in the client info string to show how much memory is used up
by buffered multi commands.
* Client `tot-mem` now accounts for buffered multi-commands, pubsub patterns and
channels (partially), tracking prefixes (partially).
* CLIENT_CLOSE_ASAP flag is now handled in a new `beforeNextClient()` function so
clients will be disconnected between processing different clients and not only before sleep.
This new function can be used in the future for work we want to do outside the command
processing loop but don't want to wait for all clients to be processed before we get to it.
Specifically I wanted to handle output-buffer-limit related closing before we process client
eviction in case the two race with each other.
* Added a `DEBUG CLIENT-EVICTION` command to print out info about the client eviction
buckets.
* Each client now holds a pointer to the client eviction memory usage bucket it belongs to
and listNode to itself in that bucket for quick removal.
* Global `io_threads_op` variable now can contain a `IO_THREADS_OP_IDLE` value
indicating no io-threading is currently being executed.
* In order to track memory used by each clients in real-time we can't rely on updating
these stats in `clientsCron()` alone anymore. So now I call `updateClientMemUsage()`
(used to be `clientsCronTrackClientsMemUsage()`) after command processing, after
writing data to pubsub clients, after writing the output buffer and after reading from the
socket (and maybe other places too). The function is written to be fast.
* Clients are evicted if needed (with appropriate log line) in `beforeSleep()` and before
processing a command (before performing oom-checks and key-eviction).
* All clients memory usage buckets are grouped as follows:
* All clients using less than 64k.
* 64K..128K
* 128K..256K
* ...
* 2G..4G
* All clients using 4g and up.
* Added client-eviction.tcl with a bunch of tests for the new mechanism.
* Extended maxmemory.tcl to test the interaction between maxmemory and
maxmemory-clients settings.
* Added an option to flag a numeric configuration variable as a "percent", this means that
if we encounter a '%' after the number in the config file (or config set command) we
consider it as valid. Such a number is store internally as a negative value. This way an
integer value can be interpreted as either a percent (negative) or absolute value (positive).
This is useful for example if some numeric configuration can optionally be set to a percentage
of something else.
Co-authored-by: Oran Agra <oran@redislabs.com>
* Delay to discard cache master when full synchronization
* Don't disconnect with replicas before loading transferred RDB when full sync
Previously, once replica need to start full synchronization with master,
it will discard cached master whatever full synchronization is failed or
not.
Now we discard cached master only when transferring RDB is finished
and start to change data space, this make replica could start partial
resynchronization with another new master if new master is failed
during full synchronization.
Reduce dict struct memory overhead
on 64bit dict size goes down from jemalloc's 96 byte bin to its 56 byte bin.
summary of changes:
- Remove `privdata` from callbacks and dict creation. (this affects many files, see "Interface change" below).
- Meld `dictht` struct into the `dict` struct to eliminate struct padding. (this affects just dict.c and defrag.c)
- Eliminate the `sizemask` field, can be calculated from size when needed.
- Convert the `size` field into `size_exp` (exponent), utilizes one byte instead of 8.
Interface change: pass dict pointer to dict type call back functions.
This is instead of passing the removed privdata field. In the future if
we'd like to have private data in the callbacks we can extract it from
the dict type. We can extend dictType to include a custom dict struct
allocator and use it to allocate more data at the end of the dict
struct. This data can then be used to store private data later acccessed
by the callbacks.
1. In sendBulkToSlave, we used LL_VERBOSE in the past, changed to
LL_WARNING. (all the other places that do freeClient(slave) use LL_WARNING)
2. The old style LOG_WARNING, chang it to LL_WARNING. Introduced in an
old pr (#1690).
in the past, the reply list was a list of sds objects, so this didn't have any overhead,
but now addReplySds just copies the data from the sds and frees it, so there's no
need to make a copy of the buffer before copying again.
this reduces an excessive allocation and free and a memcpy.
In the past, the first bind address that was explicitly specified was
also used to bind outgoing connections. This could result with some
problems. For example: on some systems using `bind 127.0.0.1` would
result with outgoing connections also binding to `127.0.0.1` and failing
to connect to remote addresses.
With the recent change to the way `bind` is handled, this presented
other issues:
* The default first bind address is '*' which is not a valid address.
* We make no distinction between user-supplied config that is identical
to the default, and the default config.
This commit addresses both these issues by introducing an explicit
configuration parameter to control the bind address on outgoing
connections.
This PR adds a spell checker CI action that will fail future PRs if they introduce typos and spelling mistakes.
This spell checker is based on blacklist of common spelling mistakes, so it will not catch everything,
but at least it is also unlikely to cause false positives.
Besides that, the PR also fixes many spelling mistakes and types, not all are a result of the spell checker we use.
Here's a summary of other changes:
1. Scanned the entire source code and fixes all sorts of typos and spelling mistakes (including missing or extra spaces).
2. Outdated function / variable / argument names in comments
3. Fix outdated keyspace masks error log when we check `config.notify-keyspace-events` in loadServerConfigFromString.
4. Trim the white space at the end of line in `module.c`. Check: https://github.com/redis/redis/pull/7751
5. Some outdated https link URLs.
6. Fix some outdated comment. Such as:
- In README: about the rdb, we used to said create a `thread`, change to `process`
- dbRandomKey function coment (about the dictGetRandomKey, change to dictGetFairRandomKey)
- notifyKeyspaceEvent fucntion comment (add type arg)
- Some others minor fix in comment (Most of them are incorrectly quoted by variable names)
7. Modified the error log so that users can easily distinguish between TCP and TLS in `changeBindAddr`
When replica never successfully connect to master, server.repl_down_since
will be initialized to 0, therefore, the info master_link_down_since_seconds
was showing the current unix timestamp, which does not make much sense.
This commit fixes the issue by showing master_link_down_since_seconds to -1.
means the replica never connect to master before.
This commit also resets this variable back to 0 when a replica is turned into
a master, so that it'll behave the same if the master is later turned into a
replica again.
The implication of this change is that if some app is checking if the value is > 60
do something, like conclude the replica is stale, this could case harm (changing
a big positive number with a small one).
Starting redis 6.0 (part of the TLS feature), diskless master uses pipe from the fork
child so that the parent is the one sending data to the replicas.
This mechanism has an issue in which a hung replica will cause the master to wait
for it to read the data sent to it forever, thus preventing the fork child from terminating
and preventing the creations of any other forks.
This PR adds a timeout mechanism, much like the ACK-based timeout,
we disconnect replicas that aren't reading the RDB file fast enough.
On a replica we do accept connections, even though commands accessing
the database will operate in read-only mode. But the server is still
already operational and processing commands.
Not sending the readiness notification means that on a HA setup where
the nodes all start as replicas (with replicaof in the config) with
a replica that cannot connect to the master server and which might not
come back in a predictable amount of time or at all, the service
supervisor will end up timing out the service and terminating it, with
no option to promote it to be the main instance. This seems counter to
what the readiness notification is supposed to be signaling.
Instead send the readiness notification when we start accepting
commands, and then send the various server status changes as that.
Fixes: commit 641c64ada1
Fixes: commit dfb598cf33
Originally this was limited to IPv6 address length, but effectively it
has been used for host names and now that Sentinel accepts that as well
we need to be able to store full hostnames.
Fixes#8507
In some scenarios, such as remote backup, we only want to get remote
redis server db snapshot. Currently, redis-cli acts as a replica and
sends SYNC to redis, but redis still accumulates replication buffer
in the replica client output buffer, that may result in using vast
memory, or failing to transfer RDB because of client-output-buffer-limit.
In this commit, we add 'replconf rdb-only 0|1', redis doesn't send
incremental replication buffer to them if they send 'replconf rdb-only 1',
so we can reduce used memory and improve success of getting RDB.
This is a refactory commit, isn't suppose to have any actual impact.
it does the following:
- keep just one server struct fork child pid variable instead of 3
- have one server struct variable indicating the purpose of the current fork
child.
- redisFork is now responsible of updating the server struct with the pid,
which means it can be the one that calls updateDictResizePolicy
- move child info pipe handling into redisFork instead of having them
repeated outside
- there are two classes of fork purposes, mutually exclusive group (AOF, RDB,
Module), and one that can create several forks to coexist in parallel (LDB,
but maybe Modules some day too, Module API allows for that).
- minor fix to killRDBChild:
unlike killAppendOnlyChild and TerminateModuleForkChild, the killRDBChild
doesn't clear the pid variable or call wait4, so checkChildrenDone does
the cleanup for it.
This commit removes the explicit calls to rdbRemoveTempFile, closeChildInfoPipe,
updateDictResizePolicy, which didn't do any harm, but where unnecessary.
In #7726 (part of 6.2), we added a mechanism for whileBlockedCron, this
mechanism has an assertion to make sure the timestamp in
whileBlockedCron was always set correctly before the blocking operation
starts.
I now found (thanks to our CI) two bugs in that area:
1) CONFIG RESETSTAT (if it was allowed during loading) would have
cleared this var
2) the call stopLoading (which calls whileBlockedCron) was made too
early, while the rio is still in use, in which case the update_cksum
(rdbLoadProgressCallback) may still be called and whileBlockedCron
can assert.
The commit deals with the syncWithMaster and the ugly state machine in it.
It attempts to clean it a bit, but more importantly it uses pipeline for
part of the work (rather than 7 round trips, we now have 4).
i.e. the connect and PING are separate, then AUTH + 3 REPLCONF in one pipeline,
and finally the PSYNC (must be separate since the master has to have an empty
output buffer).
This is just a refactoring commit.
This function was never actually used as a synchronous (do both send or
receive), it was always used only ine one of the two modes, which meant it
has to take extra arguments that are not relevant for the other.
Besides that, a tool that sends a synchronous command, it not something
we want in our toolbox (synchronous IO in single threaded app is evil).
sendSynchronousCommand was now refactored into separate sending and
receiving APIs, and the sending part has two variants, one taking vaargs,
and the other taking argc+argv (and an optional length array which means
you can use binary sds strings).
When replica diskless-load type is swapdb in cluster mode, we didn't backup
keys to slots map, so we will lose keys to slots map if fail to sync.
Now we backup keys to slots map at first, and restore it properly when fail.
This commit includes a refactory/cleanup of the backups mechanism (moving it to db.c and re-structuring it a bit).
Co-authored-by: Oran Agra <oran@redislabs.com>
If we enable diskless replication, set repl-diskless-sync-delay to 0,
and master has non-rdb child process such as rewrite aof child, master
will try to start to a new BGSAVE but fails immediately (before fork)
when replicas ask for full synchronization, and master always fails
to start a new BGSAVE and disconnects with replicas until non-rdb
child process exists.
this bug was introduced in #6271 (not yet released in 6.0.x)
Blocking command should not be used with MULTI, LUA, and RM_Call. This is because,
the caller, who executes the command in this context, expects a reply.
Today, LUA and MULTI have a special (and different) treatment to blocking commands:
LUA - Most commands are marked with no-script flag which are checked when executing
and command from LUA, commands that are not marked (like XREAD) verify that their
blocking mode is not used inside LUA (by checking the CLIENT_LUA client flag).
MULTI - Command that is going to block, first verify that the client is not inside
multi (by checking the CLIENT_MULTI client flag). If the client is inside multi, they
return a result which is a match to the empty key with no timeout (for example blpop
inside MULTI will act as lpop)
For modules that perform RM_Call with blocking command, the returned results type is
REDISMODULE_REPLY_UNKNOWN and the caller can not really know what happened.
Disadvantages of the current state are:
No unified approach, LUA, MULTI, and RM_Call, each has a different treatment
Module can not safely execute blocking command (and get reply or error).
Though It is true that modules are not like LUA or MULTI and should be smarter not
to execute blocking commands on RM_Call, sometimes you want to execute a command base
on client input (for example if you create a module that provides a new scripting
language like javascript or python).
While modules (on modules command) can check for REDISMODULE_CTX_FLAGS_LUA or
REDISMODULE_CTX_FLAGS_MULTI to know not to block the client, there is no way to
check if the command came from another module using RM_Call. So there is no way
for a module to know not to block another module RM_Call execution.
This commit adds a way to unify the treatment for blocking clients by introducing
a new CLIENT_DENY_BLOCKING client flag. On LUA, MULTI, and RM_Call the new flag
turned on to signify that the client should not be blocked. A blocking command
verifies that the flag is turned off before blocking. If a blocking command sees
that the CLIENT_DENY_BLOCKING flag is on, it's not blocking and return results
which are matches to empty key with no timeout (as MULTI does today).
The new flag is checked on the following commands:
List blocking commands: BLPOP, BRPOP, BRPOPLPUSH, BLMOVE,
Zset blocking commands: BZPOPMIN, BZPOPMAX
Stream blocking commands: XREAD, XREADGROUP
SUBSCRIBE, PSUBSCRIBE, MONITOR
In addition, the new flag is turned on inside the AOF client, we do not want to
block the AOF client to prevent deadlocks and commands ordering issues (and there
is also an existing assert in the code that verifies it).
To keep backward compatibility on LUA, all the no-script flags on existing commands
were kept untouched. In addition, a LUA special treatment on XREAD and XREADGROUP was kept.
To keep backward compatibility on MULTI (which today allows SUBSCRIBE, and PSUBSCRIBE).
We added a special treatment on those commands to allow executing them on MULTI.
The only backward compatibility issue that this PR introduces is that now MONITOR
is not allowed inside MULTI.
Tests were added to verify blocking commands are not blocking the client on LUA, MULTI,
or RM_Call. Tests were added to verify the module can check for CLIENT_DENY_BLOCKING flag.
Co-authored-by: Oran Agra <oran@redislabs.com>
Co-authored-by: Itamar Haber <itamar@redislabs.com>
Useful when you want to know through which bind address the client connected to
the server in case of multiple bind addresses.
- Adding `laddr` field to CLIENT list showing the local (bind) address.
- Adding `LADDR` option to CLIENT KILL to kill all the clients connected
to a specific local address.
- Refactoring to share code.
If we fail or stop to rewrite aof, we need to remove temporary aof.
We also remove temporary rdb when replicas abort to receive rdb.
But currently we delete them in main thread, to avoid blocking,
we should use bg_unlink to remove them in a background thread.
Btw, we have already used this way to removed child process temporary rdb.
We should sync temp DB file before renaming as rdb_fsync_range does not use
flag `SYNC_FILE_RANGE_WAIT_AFTER`.
Refer to `Linux Programmer's Manual`:
SYNC_FILE_RANGE_WAIT_AFTER
Wait upon write-out of all pages in the range after performing any write.
This happens only on diskless replicas when attempting to reconnect after
failing to load an RDB file. It is more likely to occur with larger datasets.
After reconnection is initiated, replicationEmptyDbCallback() may get called
and try to write to an unconnected socket. This triggered another issue where
the connection is put into an error state and the connect handler never gets
called. The problem is a regression introduced by commit c17e597.
Redis 6.0 introduces I/O threads, it is so cool and efficient, we use C11
_Atomic to establish inter-thread synchronization without mutex. But the
compiler that must supports C11 _Atomic can compile redis code, that brings a
lot of inconvenience since some common platforms can't support by default such
as CentOS7, so we want to implement redis atomic type to make it more portable.
We have implemented our atomic variable for redis that only has 'relaxed'
operations in src/atomicvar.h, so we implement some operations with
'sequentially-consistent', just like the default behavior of C11 _Atomic that
can establish inter-thread synchronization. And we replace all uses of C11
_Atomic with redis atomic variable.
Our implementation of redis atomic variable uses C11 _Atomic, __atomic or
__sync macros if available, it supports most common platforms, and we will
detect automatically which feature we use. In Makefile we use a dummy file to
detect if the compiler supports C11 _Atomic. Now for gcc, we can compile redis
code theoretically if your gcc version is not less than 4.1.2(starts to support
__sync_xxx operations). Otherwise, we remove use mutex fallback to implement
redis atomic variable for performance and test. You will get compiling errors
if your compiler doesn't support all features of above.
For cover redis atomic variable tests, we add other CI jobs that build redis on
CentOS6 and CentOS7 and workflow daily jobs that run the tests on them.
For them, we just install gcc by default in order to cover different compiler
versions, gcc is 4.4.7 by default installation on CentOS6 and 4.8.5 on CentOS7.
We restore the feature that we can test redis with Helgrind to find data race
errors. But you need install Valgrind in the default path configuration firstly
before running your tests, since we use macros in helgrind.h to tell Helgrind
inter-thread happens-before relationship explicitly for avoiding false positives.
Please open an issue on github if you find data race errors relate to this commit.
Unrelated:
- Fix redefinition of typedef 'RedisModuleUserChangedFunc'
For some old version compilers, they will report errors or warnings, if we
re-define function type.
Starting redis 6.0 and the changes we made to the diskless master to be
suitable for TLS, I made the master avoid reaping (wait3) the pid of the
child until we know all replicas are done reading their rdb.
I did that in order to avoid a state where the rdb_child_pid is -1 but
we don't yet want to start another fork (still busy serving that data to
replicas).
It turns out that the solution used so far was problematic in case the
fork child was being killed (e.g. by the kernel OOM killer), in that
case there's a chance that we currently disabled the read event on the
rdb pipe, since we're waiting for a replica to become writable again.
and in that scenario the master would have never realized the child
exited, and the replica will remain hung too.
Note that there's no mechanism to detect a hung replica while it's in
rdb transfer state.
The solution here is to add another pipe which is used by the parent to
tell the child it is safe to exit. this mean that when the child exits,
for whatever reason, it is safe to reap it.
Besides that, i'm re-introducing an adjustment to REPLCONF ACK which was
part of #6271 (Accelerate diskless master connections) but was dropped
when that PR was rebased after the TLS fork/pipe changes (5a47794).
Now that RdbPipeCleanup no longer calls checkChildrenDone, and the ACK
has chance to detect that the child exited, it should be the one to call
it so that we don't have to wait for cron (server.hz) to do that.
Client that issued WAIT last will most likely have the highest replication offset, so imagine a probably common case where all clients are waiting for the same number of replicas. we prefer the loop to start from the last client (one waiting for the highest offset), so that the optimization in the function will call replicationCountAcksByOffset for each client until it found a good one, and stop calling it for the rest of the clients.
the way the loop was implemented would mean that in such case it is likely to call replicationCountAcksByOffset for all clients.
Note: the change from > to >= is not directly related to the above.
Co-authored-by: 曹正斌 <caozb@jiedaibao.com>
Add Linux kernel OOM killer control option.
This adds the ability to control the Linux OOM killer oom_score_adj
parameter for all Redis processes, depending on the process role (i.e.
master, replica, background child).
A oom-score-adj global boolean flag control this feature. In addition,
specific values can be configured using oom-score-adj-values if
additional tuning is required.
Diskless master has some inherent latencies.
1) fork starts with delay from cron rather than immediately
2) replica is put online only after an ACK. but the ACK
was sent only once a second.
3) but even if it would arrive immediately, it will not
register in case cron didn't yet detect that the fork is done.
Besides that, when a replica disconnects, it doesn't immediately
attempts to re-connect, it waits for replication cron (one per second).
in case it was already online, it may be important to try to re-connect
as soon as possible, so that the backlog at the master doesn't vanish.
In case it disconnected during rdb transfer, one can argue that it's
not very important to re-connect immediately, but this is needed for the
"diskless loading short read" test to be able to run 100 iterations in 5
seconds, rather than 3 (waiting for replication cron re-connection)
changes in this commit:
1) sync command starts a fork immediately if no sync_delay is configured
2) replica sends REPLCONF ACK when done reading the rdb (rather than on 1s cron)
3) when a replica unexpectedly disconnets, it immediately tries to
re-connect rather than waiting 1s
4) when when a child exits, if there is another replica waiting, we spawn a new
one right away, instead of waiting for 1s replicationCron.
5) added a call to connectWithMaster from replicationSetMaster. which is called
from the REPLICAOF command but also in 3 places in cluster.c, in all of
these the connection attempt will now be immediate instead of delayed by 1
second.
side note:
we can add a call to rdbPipeReadHandler in replconfCommand when getting
a REPLCONF ACK from the replica to solve a race where the replica got
the entire rdb and EOF marker before we detected that the pipe was
closed.
in the test i did see this race happens in one about of some 300 runs,
but i concluded that this race is unlikely in real life (where the
replica is on another host and we're more likely to first detect the
pipe was closed.
the test runs 100 iterations in 3 seconds, so in some cases it'll take 4
seconds instead (waiting for another REPLCONF ACK).
Removing unneeded startBgsaveForReplication from updateSlavesWaitingForBgsave
Now that CheckChildrenDone is calling the new replicationStartPendingFork
(extracted from serverCron) there's actually no need to call
startBgsaveForReplication from updateSlavesWaitingForBgsave anymore,
since as soon as updateSlavesWaitingForBgsave returns, CheckChildrenDone is
calling replicationStartPendingFork that handles that anyway.
The code in updateSlavesWaitingForBgsave had a bug in which it ignored
repl-diskless-sync-delay, but removing that code shows that this bug was
hiding another bug, which is that the max_idle should have used >= and
not >, this one second delay has a big impact on my new test.
in case the rdb child failed, crashed or terminated unexpectedly redis
would have marked the replica clients with repl_put_online_on_ack and
then kill them only after a minute when no ack was received.
it would not stream anything to these connections, so the only effect of
this bug is a delay of 1 minute in the replicas attempt to re-connect.
