## Set replica-lazy-flush and lazyfree-lazy-user-flush to yes by
default.
There are many problems with running flush synchronously. Even in
single CPU environments, the thread managers should balance between
the freeing and serving incoming requests.
## Set lazy eviction, expire, server-del, user-del to yes by default
We now have a del and a lazyfree del, we also have these configuration
items to control: lazyfree-lazy-eviction, lazyfree-lazy-expire,
lazyfree-lazy-server-del, lazyfree-lazy-user-del. In most cases lazyfree
is better since it reduces the risk of blocking the main thread, and
because we have lazyfreeGetFreeEffort, on those with high effor
(currently
64) will use lazyfree.
Part of #653.
---------
Signed-off-by: Binbin <binloveplay1314@qq.com>
Fix feedback loop in key eviction with tracking clients when using I/O
threads.
Current issue:
Evicting keys while tracking clients or key space-notification exist
creates a feedback loop when using I/O threads:
While evicting keys we send tracking async writes to I/O threads,
preventing immediate release of tracking clients' COB memory
consumption.
Before the I/O thread finishes its write, we recheck used_memory, which
now includes the tracking clients' COB and thus continue to evict more
keys.
**Fix:**
We will skip the test for now while IO threads are active. We may
consider avoiding sending writes in `processPendingWrites` to I/O
threads for tracking clients when we are out of memory.
---------
Signed-off-by: Uri Yagelnik <uriy@amazon.com>
Signed-off-by: Madelyn Olson <madelyneolson@gmail.com>
Co-authored-by: Madelyn Olson <madelyneolson@gmail.com>
This PR is 1 of 3 PRs intended to achieve the goal of 1 million requests
per second, as detailed by [dan touitou](https://github.com/touitou-dan)
in https://github.com/valkey-io/valkey/issues/22. This PR modifies the
IO threads to be fully asynchronous, which is a first and necessary step
to allow more work offloading and better utilization of the IO threads.
### Current IO threads state:
Valkey IO threads were introduced in Redis 6.0 to allow better
utilization of multi-core machines. Before this, Redis was
single-threaded and could only use one CPU core for network and command
processing. The introduction of IO threads helps in offloading the IO
operations to multiple threads.
**Current IO Threads flow:**
1. Initialization: When Redis starts, it initializes a specified number
of IO threads. These threads are in addition to the main thread, each
thread starts with an empty list, the main thread will populate that
list in each event-loop with pending-read-clients or
pending-write-clients.
2. Read Phase: The main thread accepts incoming connections and reads
requests from clients. The reading of requests are offloaded to IO
threads. The main thread puts the clients ready-to-read in a list and
set the global io_threads_op to IO_THREADS_OP_READ, the IO threads pick
the clients up, perform the read operation and parse the first incoming
command.
3. Command Processing: After reading the requests, command processing is
still single-threaded and handled by the main thread.
4. Write Phase: Similar to the read phase, the write phase is also be
offloaded to IO threads. The main thread prepares the response in the
clients’ output buffer then the main thread puts the client in the list,
and sets the global io_threads_op to the IO_THREADS_OP_WRITE. The IO
threads then pick the clients up and perform the write operation to send
the responses back to clients.
5. Synchronization: The main-thread communicate with the threads on how
many jobs left per each thread with atomic counter. The main-thread
doesn’t access the clients while being handled by the IO threads.
**Issues with current implementation:**
* Underutilized Cores: The current implementation of IO-threads leads to
the underutilization of CPU cores.
* The main thread remains responsible for a significant portion of
IO-related tasks that could be offloaded to IO-threads.
* When the main-thread is processing client’s commands, the IO threads
are idle for a considerable amount of time.
* Notably, the main thread's performance during the IO-related tasks is
constrained by the speed of the slowest IO-thread.
* Limited Offloading: Currently, Since the Main-threads waits
synchronously for the IO threads, the Threads perform only read-parse,
and write operations, with parsing done only for the first command. If
the threads can do work asynchronously we may offload more work to the
threads reducing the load from the main-thread.
* TLS: Currently, we don't support IO threads with TLS (where offloading
IO would be more beneficial) since TLS read/write operations are not
thread-safe with the current implementation.
### Suggested change
Non-blocking main thread - The main thread and IO threads will operate
in parallel to maximize efficiency. The main thread will not be blocked
by IO operations. It will continue to process commands independently of
the IO thread's activities.
**Implementation details**
**Inter-thread communication.**
* We use a static, lock-free ring buffer of fixed size (2048 jobs) for
the main thread to send jobs and for the IO to receive them. If the ring
buffer fills up, the main thread will handle the task itself, acting as
back pressure (in case IO operations are more expensive than command
processing). A static ring buffer is a better candidate than a dynamic
job queue as it eliminates the need for allocation/freeing per job.
* An IO job will be in the format: ` [void* function-call-back | void
*data] `where data is either a client to read/write from and the
function-ptr is the function to be called with the data for example
readQueryFromClient using this format we can use it later to offload
other types of works to the IO threads.
* The Ring buffer is one way from the main-thread to the IO thread, Upon
read/write event the main thread will send a read/write job then in
before sleep it will iterate over the pending read/write clients to
checking for each client if the IO threads has already finished handling
it. The IO thread signals it has finished handling a client read/write
by toggling an atomic flag read_state / write_state on the client
struct.
**Thread Safety**
As suggested in this solution, the IO threads are reading from and
writing to the clients' buffers while the main thread may access those
clients.
We must ensure no race conditions or unsafe access occurs while keeping
the Valkey code simple and lock free.
Minimal Action in the IO Threads
The main change is to limit the IO thread operations to the bare
minimum. The IO thread will access only the client's struct and only the
necessary fields in this struct.
The IO threads will be responsible for the following:
* Read Operation: The IO thread will only read and parse a single
command. It will not update the server stats, handle read errors, or
parsing errors. These tasks will be taken care of by the main thread.
* Write Operation: The IO thread will only write the available data. It
will not free the client's replies, handle write errors, or update the
server statistics.
To achieve this without code duplication, the read/write code has been
refactored into smaller, independent components:
* Functions that perform only the read/parse/write calls.
* Functions that handle the read/parse/write results.
This refactor accounts for the majority of the modifications in this PR.
**Client Struct Safe Access**
As we ensure that the IO threads access memory only within the client
struct, we need to ensure thread safety only for the client's struct's
shared fields.
* Query Buffer
* Command parsing - The main thread will not try to parse a command from
the query buffer when a client is offloaded to the IO thread.
* Client's memory checks in client-cron - The main thread will not
access the client query buffer if it is offloaded and will handle the
querybuf grow/shrink when the client is back.
* CLIENT LIST command - The main thread will busy-wait for the IO thread
to finish handling the client, falling back to the current behavior
where the main thread waits for the IO thread to finish their
processing.
* Output Buffer
* The IO thread will not change the client's bufpos and won't free the
client's reply lists. These actions will be done by the main thread on
the client's return from the IO thread.
* bufpos / block→used: As the main thread may change the bufpos, the
reply-block→used, or add/delete blocks to the reply list while the IO
thread writes, we add two fields to the client struct: io_last_bufpos
and io_last_reply_block. The IO thread will write until the
io_last_bufpos, which was set by the main-thread before sending the
client to the IO thread. If more data has been added to the cob in
between, it will be written in the next write-job. In addition, the main
thread will not trim or merge reply blocks while the client is
offloaded.
* Parsing Fields
* Client's cmd, argc, argv, reqtype, etc., are set during parsing.
* The main thread will indicate to the IO thread not to parse a cmd if
the client is not reset. In this case, the IO thread will only read from
the network and won't attempt to parse a new command.
* The main thread won't access the c→cmd/c→argv in the CLIENT LIST
command as stated before it will busy wait for the IO threads.
* Client Flags
* c→flags, which may be changed by the main thread in multiple places,
won't be accessed by the IO thread. Instead, the main thread will set
the c→io_flags with the information necessary for the IO thread to know
the client's state.
* Client Close
* On freeClient, the main thread will busy wait for the IO thread to
finish processing the client's read/write before proceeding to free the
client.
* Client's Memory Limits
* The IO thread won't handle the qb/cob limits. In case a client crosses
the qb limit, the IO thread will stop reading for it, letting the main
thread know that the client crossed the limit.
**TLS**
TLS is currently not supported with IO threads for the following
reasons:
1. Pending reads - If SSL has pending data that has already been read
from the socket, there is a risk of not calling the read handler again.
To handle this, a list is used to hold the pending clients. With IO
threads, multiple threads can access the list concurrently.
2. Event loop modification - Currently, the TLS code
registers/unregisters the file descriptor from the event loop depending
on the read/write results. With IO threads, multiple threads can modify
the event loop struct simultaneously.
3. The same client can be sent to 2 different threads concurrently
(https://github.com/redis/redis/issues/12540).
Those issues were handled in the current PR:
1. The IO thread only performs the read operation. The main thread will
check for pending reads after the client returns from the IO thread and
will be the only one to access the pending list.
2. The registering/unregistering of events will be similarly postponed
and handled by the main thread only.
3. Each client is being sent to the same dedicated thread (c→id %
num_of_threads).
**Sending Replies Immediately with IO threads.**
Currently, after processing a command, we add the client to the
pending_writes_list. Only after processing all the clients do we send
all the replies. Since the IO threads are now working asynchronously, we
can send the reply immediately after processing the client’s requests,
reducing the command latency. However, if we are using AOF=always, we
must wait for the AOF buffer to be written, in which case we revert to
the current behavior.
**IO threads dynamic adjustment**
Currently, we use an all-or-nothing approach when activating the IO
threads. The current logic is as follows: if the number of pending write
clients is greater than twice the number of threads (including the main
thread), we enable all threads; otherwise, we enable none. For example,
if 8 IO threads are defined, we enable all 8 threads if there are 16
pending clients; else, we enable none.
It makes more sense to enable partial activation of the IO threads. If
we have 10 pending clients, we will enable 5 threads, and so on. This
approach allows for a more granular and efficient allocation of
resources based on the current workload.
In addition, the user will now be able to change the number of I/O
threads at runtime. For example, when decreasing the number of threads
from 4 to 2, threads 3 and 4 will be closed after flushing their job
queues.
**Tests**
Currently, we run the io-threads tests with 4 IO threads
(443d80f168/.github/workflows/daily.yml (L353)).
This means that we will not activate the IO threads unless there are 8
(threads * 2) pending write clients per single loop, which is unlikely
to happened in most of tests, meaning the IO threads are not currently
being tested.
To enforce the main thread to always offload work to the IO threads,
regardless of the number of pending events, we add an
events-per-io-thread configuration with a default value of 2. When set
to 0, this configuration will force the main thread to always offload
work to the IO threads.
When we offload every single read/write operation to the IO threads, the
IO-threads are running with 100% CPU when running multiple tests
concurrently some tests fail as a result of larger than expected command
latencies. To address this issue, we have to add some after or wait_for
calls to some of the tests to ensure they pass with IO threads as well.
Signed-off-by: Uri Yagelnik <uriy@amazon.com>
When LRU/LFU enabled, Valkey does not allow using shared objects, as
value objects may be shared among many different keys and they can't
share LRU/LFU information.
However `maxmemory-policy` is modifiable at runtime. If LRU/LFU is not
enabled at start, but then enabled when some shared objects are already
used, there could be some confusion in LRU/LFU information.
For `set` command it is OK since it is going to create a new object when
LRU/LFU enabled, but `get` command will not unshare the object and just
update LRU/LFU information.
So we may duplicate the object in this case. It is a one-time task for
each key using shared objects, unless this is the case for so many keys,
there should be no serious performance degradation.
Still, LRU will be updated anyway, no matter LRU/LFU is enabled or not,
because `OBJECT IDLETIME` needs it, unless `maxmemory-policy` is set to
LFU. So idle time of a key may still be messed up.
---------
Signed-off-by: chentianjie.ctj <chentianjie.ctj@alibaba-inc.com>
Signed-off-by: Chen Tianjie <TJ_Chen@outlook.com>
Updated procedure redis_deferring_client in test environent to
valkey_deferring_client.
Signed-off-by: Shivshankar-Reddy <shiva.sheri.github@gmail.com>
This includes comments used for module API documentation.
* Strategy for replacement: Regex search: `(//|/\*| \*|#).* ("|\()?(r|R)edis( |\.
|'|\n|,|-|\)|")(?!nor the names of its contributors)(?!Ltd.)(?!Labs)(?!Contributors.)`
* Don't edit copyright comments
* Replace "Redis version X.X" -> "Redis OSS version X.X" to distinguish
from newly licensed repository
* Replace "Redis Object" -> "Object"
* Exclude markdown for now
* Don't edit Lua scripting comments referring to redis.X API
* Replace "Redis Protocol" -> "RESP"
* Replace redis-benchmark, -cli, -server, -check-aof/rdb with "valkey-"
prefix
* Most other places, I use best judgement to either remove "Redis", or
replace with "the server" or "server"
Fixes#148
---------
Signed-off-by: Jacob Murphy <jkmurphy@google.com>
Signed-off-by: Viktor Söderqvist <viktor.soderqvist@est.tech>
Ci report this failure:
```
*** [err]: Don't rehash if used memory exceeds maxmemory after rehash in tests/unit/maxmemory.tcl
Expected '4098' to equal or match '4002'
WARNING: the new maxmemory value set via CONFIG SET (1176088) is smaller than the current memory usage (1231083)
```
It can be seen from the log that used_memory changed before we set
maxmemory.
The reason is that in #12819, in cron, in addition to trying to shrink,
we will
also tyring to expand. The dict was expanded by cron before we set
maxmemory,
causing the test to fail.
Before setting maxmemory, we only add 4095 keys to avoid triggering
resize.
This pr can get two performance benefits:
1. Stop redundant initialization when most robj objects are created
2. LRU_CLOCK will no longer be called in io threads, so we can avoid the `atomicGet`
Another code optimization:
deleted the redundant judgment in dbSetValue, no matter in LFU or LRU, the lru field inold
robj is always the freshest (it is always updated in lookupkey), so we don't need to judge if in LFU
The MacOS CI in github actions often hangs without any logs. GH argues that
it's due to resource utilization, either running out of disk space, memory, or CPU
starvation, and thus the runner is terminated.
This PR contains multiple attempts to resolve this:
1. introducing pause_process instead of SIGSTOP, which waits for the process
to stop before resuming the test, possibly resolving race conditions in some tests,
this was a suspect since there was one test that could result in an infinite loop in that
case, in practice this didn't help, but still a good idea to keep.
2. disable the `save` config in many tests that don't need it, specifically ones that use
heavy writes and could create large files.
3. change the `populate` proc to use short pipeline rather than an infinite one.
4. use `--clients 1` in the macos CI so that we don't risk running multiple resource
demanding tests in parallel.
5. enable `--verbose` to be repeated to elevate verbosity and print more info to stdout
when a test or a server starts.
This PR add `assert_refcount_morethan`, and modify `assert_refcount` to skip
the `OBJECT REFCOUNT` check with `needs:debug` flag. Use them to modify all
`OBJECT REFCOUNT` calls and also update the tests/README to be more specific.
The reasoning is that some of these tests could be testing something important,
and along the way also add a check for the refcount, and it could be a shame to skip
the whole test just because the refcount functionality is missing or blocked.
but much like the fact that some redis variants may not support DEBUG,
and still we want to run the majority of the test for coverage, and just skip the digest match.
Fix replication inconsistency on modules that uses key space notifications.
### The Problem
In general, key space notifications are invoked after the command logic was
executed (this is not always the case, we will discuss later about specific
command that do not follow this rules). For example, the `set x 1` will trigger
a `set` notification that will be invoked after the `set` logic was performed, so
if the notification logic will try to fetch `x`, it will see the new data that was written.
Consider the scenario on which the notification logic performs some write
commands. for example, the notification logic increase some counter,
`incr x{counter}`, indicating how many times `x` was changed.
The logical order by which the logic was executed is has follow:
```
set x 1
incr x{counter}
```
The issue is that the `set x 1` command is added to the replication buffer
at the end of the command invocation (specifically after the key space
notification logic was invoked and performed the `incr` command).
The replication/aof sees the commands in the wrong order:
```
incr x{counter}
set x 1
```
In this specific example the order is less important.
But if, for example, the notification would have deleted `x` then we would
end up with primary-replica inconsistency.
### The Solution
Put the command that cause the notification in its rightful place. In the
above example, the `set x 1` command logic was executed before the
notification logic, so it should be added to the replication buffer before
the commands that is invoked by the notification logic. To achieve this,
without a major code refactoring, we save a placeholder in the replication
buffer, when finishing invoking the command logic we check if the command
need to be replicated, and if it does, we use the placeholder to add it to the
replication buffer instead of appending it to the end.
To be efficient and not allocating memory on each command to save the
placeholder, the replication buffer array was modified to reuse memory
(instead of allocating it each time we want to replicate commands).
Also, to avoid saving a placeholder when not needed, we do it only for
WRITE or MAY_REPLICATE commands.
#### Additional Fixes
* Expire and Eviction notifications:
* Expire/Eviction logical order was to first perform the Expire/Eviction
and then the notification logic. The replication buffer got this in the
other way around (first notification effect and then the `del` command).
The PR fixes this issue.
* The notification effect and the `del` command was not wrap with
`multi-exec` (if needed). The PR also fix this issue.
* SPOP command:
* On spop, the `spop` notification was fired before the command logic
was executed. The change in this PR would have cause the replication
order to be change (first `spop` command and then notification `logic`)
although the logical order is first the notification logic and then the
`spop` logic. The right fix would have been to move the notification to
be fired after the command was executed (like all the other commands),
but this can be considered a breaking change. To overcome this, the PR
keeps the current behavior and changes the `spop` code to keep the right
logical order when pushing commands to the replication buffer. Another PR
will follow to fix the SPOP properly and match it to the other command (we
split it to 2 separate PR's so it will be easy to cherry-pick this PR to 7.0 if
we chose to).
#### Unhanded Known Limitations
* key miss event:
* On key miss event, if a module performed some write command on the
event (using `RM_Call`), the `dirty` counter would increase and the read
command that cause the key miss event would be replicated to the replication
and aof. This problem can also happened on a write command that open
some keys but eventually decides not to perform any action. We decided
not to handle this problem on this PR because the solution is complex
and will cause additional risks in case we will want to cherry-pick this PR.
We should decide if we want to handle it in future PR's. For now, modules
writers is advice not to perform any write commands on key miss event.
#### Testing
* We already have tests to cover cases where a notification is invoking write
commands that are also added to the replication buffer, the tests was modified
to verify that the replica gets the command in the correct logical order.
* Test was added to verify that `spop` behavior was kept unchanged.
* Test was added to verify key miss event behave as expected.
* Test was added to verify the changes do not break lazy expiration.
#### Additional Changes
* `propagateNow` function can accept a special dbid, -1, indicating not
to replicate `select`. We use this to replicate `multi/exec` on `propagatePendingCommands`
function. The side effect of this change is that now the `select` command
will appear inside the `multi/exec` block on the replication stream (instead of
outside of the `multi/exec` block). Tests was modified to match this new behavior.
# Short description
The Redis extended latency stats track per command latencies and enables:
- exporting the per-command percentile distribution via the `INFO LATENCYSTATS` command.
**( percentile distribution is not mergeable between cluster nodes ).**
- exporting the per-command cumulative latency distributions via the `LATENCY HISTOGRAM` command.
Using the cumulative distribution of latencies we can merge several stats from different cluster nodes
to calculate aggregate metrics .
By default, the extended latency monitoring is enabled since the overhead of keeping track of the
command latency is very small.
If you don't want to track extended latency metrics, you can easily disable it at runtime using the command:
- `CONFIG SET latency-tracking no`
By default, the exported latency percentiles are the p50, p99, and p999.
You can alter them at runtime using the command:
- `CONFIG SET latency-tracking-info-percentiles "0.0 50.0 100.0"`
## Some details:
- The total size per histogram should sit around 40 KiB. We only allocate those 40KiB when a command
was called for the first time.
- With regards to the WRITE overhead As seen below, there is no measurable overhead on the achievable
ops/sec or full latency spectrum on the client. Including also the measured redis-benchmark for unstable
vs this branch.
- We track from 1 nanosecond to 1 second ( everything above 1 second is considered +Inf )
## `INFO LATENCYSTATS` exposition format
- Format: `latency_percentiles_usec_<CMDNAME>:p0=XX,p50....`
## `LATENCY HISTOGRAM [command ...]` exposition format
Return a cumulative distribution of latencies in the format of a histogram for the specified command names.
The histogram is composed of a map of time buckets:
- Each representing a latency range, between 1 nanosecond and roughly 1 second.
- Each bucket covers twice the previous bucket's range.
- Empty buckets are not printed.
- Everything above 1 sec is considered +Inf.
- At max there will be log2(1000000000)=30 buckets
We reply a map for each command in the format:
`<command name> : { `calls`: <total command calls> , `histogram` : { <bucket 1> : latency , < bucket 2> : latency, ... } }`
Co-authored-by: Oran Agra <oran@redislabs.com>
The mess:
Some parts use alsoPropagate for late propagation, others using an immediate one (propagate()),
causing edge cases, ugly/hacky code, and the tendency for bugs
The basic idea is that all commands are propagated via alsoPropagate (i.e. added to a list) and the
top-most call() is responsible for going over that list and actually propagating them (and wrapping
them in MULTI/EXEC if there's more than one command). This is done in the new function,
propagatePendingCommands.
Callers to propagatePendingCommands:
1. top-most call() (we want all nested call()s to add to the also_propagate array and just the top-most
one to propagate them) - via `afterCommand`
2. handleClientsBlockedOnKeys: it is out of call() context and it may propagate stuff - via `afterCommand`.
3. handleClientsBlockedOnKeys edge case: if the looked-up key is already expired, we will propagate the
expire but will not unblock any client so `afterCommand` isn't called. in that case, we have to propagate
the deletion explicitly.
4. cron stuff: active-expire and eviction may also propagate stuff
5. modules: the module API allows to propagate stuff from just about anywhere (timers, keyspace notifications,
threads). I could have tried to catch all the out-of-call-context places but it seemed easier to handle it in one
place: when we free the context. in the spirit of what was done in call(), only the top-most freeing of a module
context may cause propagation.
6. modules: when using a thread-safe ctx it's not clear when/if the ctx will be freed. we do know that the module
must lock the GIL before calling RM_Replicate/RM_Call so we propagate the pending commands when
releasing the GIL.
A "known limitation", which were actually a bug, was fixed because of this commit (see propagate.tcl):
When using a mix of RM_Call with `!` and RM_Replicate, the command would propagate out-of-order:
first all the commands from RM_Call, and then the ones from RM_Replicate
Another thing worth mentioning is that if, in the past, a client would issue a MULTI/EXEC with just one
write command the server would blindly propagate the MULTI/EXEC too, even though it's redundant.
not anymore.
This commit renames propagate() to propagateNow() in order to cause conflicts in pending PRs.
propagatePendingCommands is the only caller of propagateNow, which is now a static, internal helper function.
Optimizations:
1. alsoPropagate will not add stuff to also_propagate if there's no AOF and replicas
2. alsoPropagate reallocs also_propagagte exponentially, to save calls to memmove
Bugfixes:
1. CONFIG SET can create evictions, sending notifications which can cause to dirty++ with modules.
we need to prevent it from propagating to AOF/replicas
2. We need to set current_client in RM_Call. buggy scenario:
- CONFIG SET maxmemory, eviction notifications, module hook calls RM_Call
- assertion in lookupKey crashes, because current_client has CONFIG SET, which isn't CMD_WRITE
3. minor: in eviction, call propagateDeletion after notification, like active-expire and all commands
(we always send a notification before propagating the command)
## Background
For redis master, one replica uses one copy of replication buffer, that is a big waste of memory,
more replicas more waste, and allocate/free memory for every reply list also cost much.
If we set client-output-buffer-limit small and write traffic is heavy, master may disconnect with
replicas and can't finish synchronization with replica. If we set client-output-buffer-limit big,
master may be OOM when there are many replicas that separately keep much memory.
Because replication buffers of different replica client are the same, one simple idea is that
all replicas only use one replication buffer, that will effectively save memory.
Since replication backlog content is the same as replicas' output buffer, now we
can discard replication backlog memory and use global shared replication buffer
to implement replication backlog mechanism.
## Implementation
I create one global "replication buffer" which contains content of replication stream.
The structure of "replication buffer" is similar to the reply list that exists in every client.
But the node of list is `replBufBlock`, which has `id, repl_offset, refcount` fields.
```c
/* Replication buffer blocks is the list of replBufBlock.
*
* +--------------+ +--------------+ +--------------+
* | refcount = 1 | ... | refcount = 0 | ... | refcount = 2 |
* +--------------+ +--------------+ +--------------+
* | / \
* | / \
* | / \
* Repl Backlog Replia_A Replia_B
*
* Each replica or replication backlog increments only the refcount of the
* 'ref_repl_buf_node' which it points to. So when replica walks to the next
* node, it should first increase the next node's refcount, and when we trim
* the replication buffer nodes, we remove node always from the head node which
* refcount is 0. If the refcount of the head node is not 0, we must stop
* trimming and never iterate the next node. */
/* Similar with 'clientReplyBlock', it is used for shared buffers between
* all replica clients and replication backlog. */
typedef struct replBufBlock {
int refcount; /* Number of replicas or repl backlog using. */
long long id; /* The unique incremental number. */
long long repl_offset; /* Start replication offset of the block. */
size_t size, used;
char buf[];
} replBufBlock;
```
So now when we feed replication stream into replication backlog and all replicas, we only need
to feed stream into replication buffer `feedReplicationBuffer`. In this function, we set some fields of
replication backlog and replicas to references of the global replication buffer blocks. And we also
need to check replicas' output buffer limit to free if exceeding `client-output-buffer-limit`, and trim
replication backlog if exceeding `repl-backlog-size`.
When sending reply to replicas, we also need to iterate replication buffer blocks and send its
content, when totally sending one block for replica, we decrease current node count and
increase the next current node count, and then free the block which reference is 0 from the
head of replication buffer blocks.
Since now we use linked list to manage replication backlog, it may cost much time for iterating
all linked list nodes to find corresponding replication buffer node. So we create a rax tree to
store some nodes for index, but to avoid rax tree occupying too much memory, i record
one per 64 nodes for index.
Currently, to make partial resynchronization as possible as much, we always let replication
backlog as the last reference of replication buffer blocks, backlog size may exceeds our setting
if slow replicas that reference vast replication buffer blocks, and this method doesn't increase
memory usage since they share replication buffer. To avoid freezing server for freeing unreferenced
replication buffer blocks when we need to trim backlog for exceeding backlog size setting,
we trim backlog incrementally (free 64 blocks per call now), and make it faster in
`beforeSleep` (free 640 blocks).
### Other changes
- `mem_total_replication_buffers`: we add this field in INFO command, it means the total
memory of replication buffers used.
- `mem_clients_slaves`: now even replica is slow to replicate, and its output buffer memory
is not 0, but it still may be 0, since replication backlog and replicas share one global replication
buffer, only if replication buffer memory is more than the repl backlog setting size, we consider
the excess as replicas' memory. Otherwise, we think replication buffer memory is the consumption
of repl backlog.
- Key eviction
Since all replicas and replication backlog share global replication buffer, we think only the
part of exceeding backlog size the extra separate consumption of replicas.
Because we trim backlog incrementally in the background, backlog size may exceeds our
setting if slow replicas that reference vast replication buffer blocks disconnect.
To avoid massive eviction loop, we don't count the delayed freed replication backlog into
used memory even if there are no replicas, i.e. we also regard this memory as replicas's memory.
- `client-output-buffer-limit` check for replica clients
It doesn't make sense to set the replica clients output buffer limit lower than the repl-backlog-size
config (partial sync will succeed and then replica will get disconnected). Such a configuration is
ignored (the size of repl-backlog-size will be used). This doesn't have memory consumption
implications since the replica client will share the backlog buffers memory.
- Drop replication backlog after loading data if needed
We always create replication backlog if server is a master, we need it because we put DELs in
it when loading expired keys in RDB, but if RDB doesn't have replication info or there is no rdb,
it is not possible to support partial resynchronization, to avoid extra memory of replication backlog,
we drop it.
- Multi IO threads
Since all replicas and replication backlog use global replication buffer, if I/O threads are enabled,
to guarantee data accessing thread safe, we must let main thread handle sending the output buffer
to all replicas. But before, other IO threads could handle sending output buffer of all replicas.
## Other optimizations
This solution resolve some other problem:
- When replicas disconnect with master since of out of output buffer limit, releasing the output
buffer of replicas may freeze server if we set big `client-output-buffer-limit` for replicas, but now,
it doesn't cause freezing.
- This implementation may mitigate reply list copy cost time(also freezes server) when one replication
has huge reply buffer and another replica can copy buffer for full synchronization. now, we just copy
reference info, it is very light.
- If we set replication backlog size big, it also may cost much time to copy replication backlog into
replica's output buffer. But this commit eliminates this problem.
- Resizing replication backlog size doesn't empty current replication backlog content.
obuf based eviction tests run until eviction occurs instead of assuming a certain
amount of writes will fill the obuf enough for eviction to occur.
This handles the kernel buffering written data and emptying the obuf even though
no one actualy reads from it.
The tests have a new timeout of 20sec: if the test doesn't pass after 20 sec it'll fail.
Hopefully this enough for our slow CI targets.
This also eliminates the need to skip some tests in TLS.
* Reduce delay between publishes to allow less time to write the obufs.
* More subscribed clients to buffer more data per publish.
* Make sure main connection isn't evicted (it has a large qbuf).
Fixing CI test issues introduced in #8687
- valgrind warnings in readQueryFromClient when client was freed by processInputBuffer
- adding DEBUG pause-cron for tests not to be time dependent.
- skipping a test that depends on socket buffers / events not compatible with TLS
- making sure client got subscribed by not using deferring client
### 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>
The initialize memory of `querybuf` is `PROTO_IOBUF_LEN(1024*16) * 2` (due to sdsMakeRoomFor being greedy), under `jemalloc`, the allocated memory will be 40k.
This will most likely result in the `querybuf` being resized when call `clientsCronResizeQueryBuffer` unless the client requests it fast enough.
Note that this bug existed even before #7875, since the condition for resizing includes the sds headers (32k+6).
## Changes
1. Use non-greedy sdsMakeRoomFor when allocating the initial query buffer (of 16k).
1. Also use non-greedy allocation when working with BIG_ARG (we won't use that extra space anyway)
2. in case we did use a greedy allocation, read as much as we can into the buffer we got (including internal frag), to reduce system calls.
3. introduce a dedicated constant for the shrinking (same value as before)
3. Add test for querybuf.
4. improve a maxmemory test by ignoring the effect of replica query buffers (can accumulate many ACKs on slow env)
5. improve a maxmemory by disabling slowlog (it will cause slight memory growth on slow env).
This commit revives the improves the ability to run the test suite against
external servers, instead of launching and managing `redis-server` processes as
part of the test fixture.
This capability existed in the past, using the `--host` and `--port` options.
However, it was quite limited and mostly useful when running a specific tests.
Attempting to run larger chunks of the test suite experienced many issues:
* Many tests depend on being able to start and control `redis-server` themselves,
and there's no clear distinction between external server compatible and other
tests.
* Cluster mode is not supported (resulting with `CROSSSLOT` errors).
This PR cleans up many things and makes it possible to run the entire test suite
against an external server. It also provides more fine grained controls to
handle cases where the external server supports a subset of the Redis commands,
limited number of databases, cluster mode, etc.
The tests directory now contains a `README.md` file that describes how this
works.
This commit also includes additional cleanups and fixes:
* Tests can now be tagged.
* Tag-based selection is now unified across `start_server`, `tags` and `test`.
* More information is provided about skipped or ignored tests.
* Repeated patterns in tests have been extracted to common procedures, both at a
global level and on a per-test file basis.
* Cleaned up some cases where test setup was based on a previous test executing
(a major anti-pattern that repeats itself in many places).
* Cleaned up some cases where test teardown was not part of a test (in the
future we should have dedicated teardown code that executes even when tests
fail).
* Fixed some tests that were flaky running on external servers.
It seems like non-Linux sockets may be less greedy, resulting with more
transient client output buffers.
Haven't proven this but empirically when stressing this test on
non-Linux tends to exhibit increased mem_clients_normal values.
c4fdf09c0 added a test that now fails with valgrind
it fails for two resons:
1) the test samples the used memory and then limits the maxmemory to
that value, but it turns out this is not atomic and on slow machines
the background cron process that clean out old query buffers reduces
the memory so that the setting doesn't cause eviction.
2) the dbsize was tested late, after reading some invalidation messages
by that time more and more keys got evicted, partially draining the
db. this is not the focus of this fix (still a known limitation)
When client tracking is enabled signalModifiedKey can increase memory usage,
this can cause the loop in performEvictions to keep running since it was measuring
the memory usage impact of signalModifiedKey.
The section that measures the memory impact of the eviction should be just on dbDelete,
excluding keyspace notification, client tracking, and propagation to AOF and replicas.
This resolves part of the problem described in #8069
p.s. fix took 1 minute, test took about 3 hours to write.
As we know, redis may reject user's requests or evict some keys if
used memory is over maxmemory. Dictionaries expanding may make
things worse, some big dictionaries, such as main db and expires dict,
may eat huge memory at once for allocating a new big hash table and be
far more than maxmemory after expanding.
There are related issues: #4213#4583
More details, when expand dict in redis, we will allocate a new big
ht[1] that generally is double of ht[0], The size of ht[1] will be
very big if ht[0] already is big. For db dict, if we have more than
64 million keys, we need to cost 1GB for ht[1] when dict expands.
If the sum of used memory and new hash table of dict needed exceeds
maxmemory, we shouldn't allow the dict to expand. Because, if we
enable keys eviction, we still couldn't add much more keys after
eviction and rehashing, what's worse, redis will keep less keys when
redis only remains a little memory for storing new hash table instead
of users' data. Moreover users can't write data in redis if disable
keys eviction.
What this commit changed ?
Add a new member function expandAllowed for dict type, it provide a way
for caller to allow expand or not. We expose two parameters for this
function: more memory needed for expanding and dict current load factor,
users can implement a function to make a decision by them.
For main db dict and expires dict type, these dictionaries may be very
big and cost huge memory for expanding, so we implement a judgement
function: we can stop dict to expand provisionally if used memory will
be over maxmemory after dict expands, but to guarantee the performance
of redis, we still allow dict to expand if dict load factor exceeds the
safe load factor.
Add test cases to verify we don't allow main db to expand when left
memory is not enough, so that avoid keys eviction.
Other changes:
For new hash table size when expand. Before this commit, the size is
that double used of dict and later _dictNextPower. Actually we aim to
control a dict load factor between 0.5 and 1.0. Now we replace *2 with
+1, since the first check is that used >= size, the outcome of before
will usually be the same as _dictNextPower(used+1). The only case where
it'll differ is when dict_can_resize is false during fork, so that later
the _dictNextPower(used*2) will cause the dict to jump to *4 (i.e.
_dictNextPower(1025*2) will return 4096).
Fix rehash test cases due to changing algorithm of new hash table size
when expand.
solving few replication related tests race conditions which fail on slow machines
bugfix in slave buffers test: since the test is executed twice, each time with
a different commands count, the threshold for the delta can't be a constant.
Few tests had borderline thresholds that were adjusted.
The slave buffers test had two issues, preventing the slave buffer from growing:
1) the slave didn't necessarily go to sleep on time, or woke up too early,
now using SIGSTOP to make sure it goes to sleep exactly when we want.
2) the master disconnected the slave on timeout
it looks like on slow machines we're getting:
[err]: slave buffer are counted correctly in tests/unit/maxmemory.tcl
Expected condition '$slave_buf > 2*1024*1024' to be true (16914 > 2*1024*1024)
this is a result of the slave waking up too early and eating the
slave buffer before the traffic and the test ends.
A) slave buffers didn't count internal fragmentation and sds unused space,
this caused them to induce eviction although we didn't mean for it.
B) slave buffers were consuming about twice the memory of what they actually needed.
- this was mainly due to sdsMakeRoomFor growing to twice as much as needed each time
but networking.c not storing more than 16k (partially fixed recently in 237a38737).
- besides it wasn't able to store half of the new string into one buffer and the
other half into the next (so the above mentioned fix helped mainly for small items).
- lastly, the sds buffers had up to 30% internal fragmentation that was wasted,
consumed but not used.
C) inefficient performance due to starting from a small string and reallocing many times.
what i changed:
- creating dedicated buffers for reply list, counting their size with zmalloc_size
- when creating a new reply node from, preallocate it to at least 16k.
- when appending a new reply to the buffer, first fill all the unused space of the
previous node before starting a new one.
other changes:
- expose mem_not_counted_for_evict info field for the benefit of the test suite
- add a test to make sure slave buffers are counted correctly and that they don't cause eviction
