Reading the PR gave me the opportunity to better specify what the code
was doing in places where I was not immediately sure about what was
going on. Moreover I documented the structure in server.h so that people
reading the header file will immediately understand what the structure
is useful for.
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
With such information will be able to use a private localtime()
implementation serverLog(), which does not use any locking and is both
thread and fork() safe.
RESTORE now supports:
1. Setting LRU/LFU
2. Absolute-time TTL
Other related changes:
1. RDB loading will not override LRU bits when RDB file
does not contain the LRU opcode.
2. RDB loading will not set LRU/LFU bits if the server's
maxmemory-policy does not match.
this reduces the extra 8 bytes we save before each pointer.
but more importantly maybe, it makes the valgrind runs to be more similiar
to our normal runs.
note: the change in malloc_stats struct in server.h is to eliminate an name conflict.
structs that are not typedefed are resolved from a separate name space.
Usually blocking operations make a lot of sense with multiple keys so
that we can listen to multiple queues (or whatever the app models) with
a single connection. However in the synchronous case it is more useful
to be able to ask for N elements. This is a change that I also wanted to
perform soon or later in the blocking list variant, but here it is more
natural since there is no reply type difference.
Implementation notes: as INFO is "already broken", I didn't want to break it further. Instead of computing the server.lua_script dict size on every call, I'm keeping a running sum of the body's length and dict overheads.
This implementation is naive as it **does not** take into consideration dict rehashing, but that inaccuracy pays off in speed ;)
Demo time:
```bash
$ redis-cli info memory | grep "script"
used_memory_scripts:96
used_memory_scripts_human:96B
number_of_cached_scripts:0
$ redis-cli eval "" 0 ; redis-cli info memory | grep "script"
(nil)
used_memory_scripts:120
used_memory_scripts_human:120B
number_of_cached_scripts:1
$ redis-cli script flush ; redis-cli info memory | grep "script"
OK
used_memory_scripts:96
used_memory_scripts_human:96B
number_of_cached_scripts:0
$ redis-cli eval "return('Hello, Script Cache :)')" 0 ; redis-cli info memory | grep "script"
"Hello, Script Cache :)"
used_memory_scripts:152
used_memory_scripts_human:152B
number_of_cached_scripts:1
$ redis-cli eval "return redis.sha1hex(\"return('Hello, Script Cache :)')\")" 0 ; redis-cli info memory | grep "script"
"1be72729d43da5114929c1260a749073732dc822"
used_memory_scripts:232
used_memory_scripts_human:232B
number_of_cached_scripts:2
✔ 19:03:54 redis [lua_scripts-in-info-memory L ✚…⚑] $ redis-cli evalsha 1be72729d43da5114929c1260a749073732dc822 0
"Hello, Script Cache :)"
```
There are too many advantages in doing this, RDB is faster to persist,
more compact, much faster to load back. The main issues here are that
the code is less tested because this was not the old default (so we are
enabling it for the new 5.0 release), and that the AOF is no longer a
trivially parsable format from now on. However the non-preamble mode
will be supported in the future as well, if new data types will be
added.
This commit, in some parts derived from PR #3041 which is no longer
possible to merge (because the user deleted the original branch),
implements the ability of slaves to have a special configuration
preventing that they try to start a failover when the master is failing.
There are multiple reasons for wanting this, and the feautre was
requested in issue #3021 time ago.
The differences between this patch and the original PR are the
following:
1. The flag is saved/loaded on the nodes configuration.
2. The 'myself' node is now flag-aware, the flag is updated as needed
when the configuration is changed via CONFIG SET.
3. The flag name uses NOFAILOVER instead of NO_FAILOVER to be consistent
with existing NOADDR.
4. The redis.conf documentation was rewritten.
Thanks to @deep011 for the original patch.
other fixes / improvements:
- LUA script memory isn't taken from zmalloc (taken from libc malloc)
so it can cause high fragmentation ratio to be displayed (which is false)
- there was a problem with "fragmentation" info being calculated from
RSS and used_memory sampled at different times (now sampling them together)
other details:
- adding a few more allocator info fields to INFO and MEMORY commands
- improve defrag test to measure defrag latency of big keys
- increasing the accuracy of the defrag test (by looking at real grag info)
this way we can use an even lower threshold and still avoid false positives
- keep the old (total) "fragmentation" field unchanged, but add new ones for spcific things
- add these the MEMORY DOCTOR command
- deduct LUA memory from the rss in case of non jemalloc allocator (one for which we don't "allocator active/used")
- reduce sampling rate of the rss and allocator info
- big keys are not defragged in one go from within the dict scan
instead they are scanned in parts after the main dict hash bucket is done.
- add latency monitor sample for defrag
- change default active-defrag-cycle-min to induce lower latency
- make active defrag start a new scan right away if needed, so it's easier
(for the test suite) to detect when it's done
- make active defrag quick the current cycle after each db / big key
- defrag some non key long term global allocations
- some refactoring for smaller functions and more reusable code
- during dict rehashing, one scan iteration of the dict, can end up scanning
one bucket in the smaller dict and many many buckets in the larger dict.
so waiting for 16 scan iterations before checking the time, may be much too long.
This commit adds two new fields in the INFO output, stats section:
expired_stale_perc:0.34
expired_time_cap_reached_count:58
The first field is an estimate of the number of keys that are yet in
memory but are already logically expired. They reason why those keys are
yet not reclaimed is because the active expire cycle can't spend more
time on the process of reclaiming the keys, and at the same time nobody
is accessing such keys. However as the active expire cycle runs, while
it will eventually have to return to the caller, because of time limit
or because there are less than 25% of keys logically expired in each
given database, it collects the stats in order to populate this INFO
field.
Note that expired_stale_perc is a running average, where the current
sample accounts for 5% and the history for 95%, so you'll see it
changing smoothly over time.
The other field, expired_time_cap_reached_count, counts the number
of times the expire cycle had to stop, even if still it was finding a
sizeable number of keys yet to expire, because of the time limit.
This allows people handling operations to understand if the Redis
server, during mass-expiration events, is able to collect keys fast
enough usually. It is normal for this field to increment during mass
expires, but normally it should very rarely increment. When instead it
constantly increments, it means that the current workloads is using
a very important percentage of CPU time to expire keys.
This feature was created thanks to the hints of Rashmi Ramesh and
Bart Robinson from Twitter. In private email exchanges, they noted how
it was important to improve the observability of this parameter in the
Redis server. Actually in big deployments, the amount of keys that are
yet to expire in each server, even if they are logically expired, may
account for a very big amount of wasted memory.
We have this operation in two places: when caching the master and
when linking a new client after the client creation. By having an API
for this we avoid incurring in errors when modifying one of the two
places forgetting the other. The function is also a good place where to
document why we cache the linked list node.
Related to #4497 and #4210.
The function in its initial form, and after the fixes for the PSYNC2
bugs, required code duplication in multiple spots. This commit modifies
it in order to always compute the script name independently, and to
return the SDS of the SHA of the body: this way it can be used in all
the places, including for SCRIPT LOAD, without duplicating the code to
create the Lua function name. Note that this requires to re-compute the
body SHA1 in the case of EVAL seeing a script for the first time, but
this should not change scripting performance in any way because new
scripts definition is a rare event happening the first time a script is
seen, and the SHA1 computation is anyway not a very slow process against
the typical Redis script and compared to the actua Lua byte compiling of
the body.
Note that the function used to assert() if a duplicated script was
loaded, however actually now two times over three, we want the function
to handle duplicated scripts just fine: this happens in SCRIPT LOAD and
in RDB AUX "lua" loading. Moreover the assert was not defending against
some obvious failure mode, so now the function always tests against
already defined functions at start.
In the case of slaves loading the RDB from master, or in other similar
cases, the script is already defined, and the function registering the
script should not fail in the assert() call.
XADD was suboptimal in the first incarnation of the command, not being
able to accept an ID (very useufl for replication), nor options for
having capped streams.
The keyspace notification for streams was not implemented.
With lists we need to signal only on key creation, but streams can
provide data to clients listening at every new item added.
To make this slightly more efficient we now track different classes of
blocked clients to avoid signaling keys when there is nobody listening.
A typical case is when the stream is used as a time series DB and
accessed only by range with XRANGE.
This is currently needed in order to fix#4483, but this can be
useful in other contexts, so maybe later we may want to remove the
conditionals and always save/load scripts.
Note that we are using the "lua" AUX field here, in order to guarantee
backward compatibility of the RDB file. The unknown AUX fields must be
discarded by past versions of Redis.
This adds a new `addReplyHelp` helper that's used by commands
when returning a help text. The following commands have been
touched: DEBUG, OBJECT, COMMAND, PUBSUB, SCRIPT and SLOWLOG.
WIP
Fix entry command table entry for OBJECT for HELP option.
After #4472 the command may have just 2 arguments.
Improve OBJECT HELP descriptions.
See #4472.
WIP 2
WIP 3
Firstly, use access time to replace the decreas time of LFU.
For function LFUDecrAndReturn,
it should only try to get decremented counter,
not update LFU fields, we will update it in an explicit way.
And we will times halve the counter according to the times of
elapsed time than server.lfu_decay_time.
Everytime a key is accessed, we should update the LFU
including update access time, and increment the counter after
call function LFUDecrAndReturn.
If a key is overwritten, the LFU should be also updated.
Then we can use `OBJECT freq` command to get a key's frequence,
and LFUDecrAndReturn should be called in `OBJECT freq` command
in case of the key has not been accessed for a long time,
because we update the access time only when the key is read or
overwritten.
In Redis 4.0 replication, with the introduction of PSYNC2, masters and
slaves replicate commands to cascading slaves and to the replication
backlog itself in a different way compared to the past.
Masters actually replicate the effects of client commands.
Slaves just propagate what they receive from masters.
This mechanism can cause problems when the configuration of an instance
is changed from master to slave inside a transaction. For instance
we could send to a master instance the following sequence:
MULTI
SLAVEOF 127.0.0.1 0
EXEC
SLAVEOF NO ONE
Before the fixes in this commit, the MULTI command used to be propagated
into the replication backlog, however after the SLAVEOF command the
instance is a slave, so the EXEC implementation failed to also propagate
the EXEC command. When the slaves of the above instance reconnected,
they were incrementally synchronized just sending a "MULTI". This put
the master client (in the slaves) into MULTI state, breaking the
replication.
Notably even Redis Sentinel uses the above approach in order to guarantee
that configuration changes are always performed together with rewrites
of the configuration and with clients disconnection. Sentiel does:
MULTI
SLAVEOF ...
CONFIG REWRITE
CLIENT KILL TYPE normal
EXEC
So this was a really problematic issue. However even with the fix in
this commit, that will add the final EXEC to the replication stream in
case the instance was switched from master to slave during the
transaction, the result would be to increment the slave replication
offset, so a successive reconnection with the new master, will not
permit a successful partial resynchronization: no way the new master can
provide us with the backlog needed, we incremented our offset to a value
that the new master cannot have.
However the EXEC implementation waits to emit the MULTI, so that if the
commands inside the transaction actually do not need to be replicated,
no commands propagation happens at all. From multi.c:
if (!must_propagate && !(c->cmd->flags & (CMD_READONLY|CMD_ADMIN))) {
execCommandPropagateMulti(c);
must_propagate = 1;
}
The above code is already modified by this commit you are reading.
Now also ADMIN commands do not trigger the emission of MULTI. It is actually
not clear why we do not just check for CMD_WRITE... Probably I wrote it this
way in order to make the code more reliable: better to over-emit MULTI
than not emitting it in time.
So this commit should indeed fix issue #3836 (verified), however it looks
like some reconsideration of this code path is needed in the long term.
BONUS POINT: The reverse bug.
Even in a read only slave "B", in a replication setup like:
A -> B -> C
There are commands without the READONLY nor the ADMIN flag, that are also
not flagged as WRITE commands. An example is just the PING command.
So if we send B the following sequence:
MULTI
PING
SLAVEOF NO ONE
EXEC
The result will be the reverse bug, where only EXEC is emitted, but not the
previous MULTI. However this apparently does not create problems in practice
but it is yet another acknowledge of the fact some work is needed here
in order to make this code path less surprising.
Note that there are many different approaches we could follow. For instance
MULTI/EXEC blocks containing administrative commands may be allowed ONLY
if all the commands are administrative ones, otherwise they could be
denined. When allowed, the commands could simply never be replicated at all.
Issue #4084 shows how for a design error, GEORADIUS is a write command
because of the STORE option. Because of this it does not work
on readonly slaves, gets redirected to masters in Redis Cluster even
when the connection is in READONLY mode and so forth.
To break backward compatibility at this stage, with Redis 4.0 to be in
advanced RC state, is problematic for the user base. The API can be
fixed into the unstable branch soon if we'll decide to do so in order to
be more consistent, and reease Redis 5.0 with this incompatibility in
the future. This is still unclear.
However, the ability to scale GEO queries in slaves easily is too
important so this commit adds two read-only variants to the GEORADIUS
and GEORADIUSBYMEMBER command: GEORADIUS_RO and GEORADIUSBYMEMBER_RO.
The commands are exactly as the original commands, but they do not
accept the STORE and STOREDIST options.
The original RDB serialization format was not parsable without the
module loaded, becuase the structure was managed only by the module
itself. Moreover RDB is a streaming protocol in the sense that it is
both produce di an append-only fashion, and is also sometimes directly
sent to the socket (in the case of diskless replication).
The fact that modules values cannot be parsed without the relevant
module loaded is a problem in many ways: RDB checking tools must have
loaded modules even for doing things not involving the value at all,
like splitting an RDB into N RDBs by key or alike, or just checking the
RDB for sanity.
In theory module values could be just a blob of data with a prefixed
length in order for us to be able to skip it. However prefixing the values
with a length would mean one of the following:
1. To be able to write some data at a previous offset. This breaks
stremaing.
2. To bufferize values before outputting them. This breaks performances.
3. To have some chunked RDB output format. This breaks simplicity.
Moreover, the above solution, still makes module values a totally opaque
matter, with the fowllowing problems:
1. The RDB check tool can just skip the value without being able to at
least check the general structure. For datasets composed mostly of
modules values this means to just check the outer level of the RDB not
actually doing any checko on most of the data itself.
2. It is not possible to do any recovering or processing of data for which a
module no longer exists in the future, or is unknown.
So this commit implements a different solution. The modules RDB
serialization API is composed if well defined calls to store integers,
floats, doubles or strings. After this commit, the parts generated by
the module API have a one-byte prefix for each of the above emitted
parts, and there is a final EOF byte as well. So even if we don't know
exactly how to interpret a module value, we can always parse it at an
high level, check the overall structure, understand the types used to
store the information, and easily skip the whole value.
The change is backward compatible: older RDB files can be still loaded
since the new encoding has a new RDB type: MODULE_2 (of value 7).
The commit also implements the ability to check RDB files for sanity
taking advantage of the new feature.
This avoids Helgrind complaining, but we are actually not using
atomicGet() to get the unixtime value for now: too many places where it
is used and given tha time_t is word-sized it should be safe in all the
archs we support as it is.
On the other hand, Helgrind, when Redis is compiled with "make helgrind"
in order to force the __sync macros, will detect the write in
updateCachedTime() as a read (because atomic functions are used) and
will not complain about races.
This commit also includes minor refactoring of mutex initializations and
a "helgrind" target in the Makefile.
Instead of giving the module background operations just a small time to
run in the beforeSleep() function, we can have the lock released for all
the time we are blocked in the multiplexing syscall.
This bug was discovered by @kevinmcgehee and constituted a major hidden
bug in the PSYNC2 implementation, caused by the propagation from the
master of incomplete commands to slaves.
The bug had several results:
1. Borrowing from Kevin text in the issue: "Given that slaves blindly
copy over their master's input into their own replication backlog over
successive read syscalls, it's possible that with large commands or
small TCP buffers, partial commands are present in this buffer. If the
master were to fail before successfully propagating the entire command
to a slave, the slaves will never execute the partial command (since the
client is invalidated) but will copy it to replication backlog which may
relay those invalid bytes to its slaves on PSYNC2, corrupting the
backlog and possibly other valid commands that follow the failover.
Simple command boundaries aren't sufficient to capture this, either,
because in the case of a MULTI/EXEC block, if the master successfully
propagates a subset of the commands but not the EXEC, then the
transaction in the backlog becomes corrupt and could corrupt other
slaves that consume this data."
2. As identified by @yangsiran later, there is another effect of the
bug. For the same mechanism of the first problem, a slave having another
slave, could receive a full resynchronization request with an already
half-applied command in the backlog. Once the RDB is ready, it will be
sent to the slave, and the replication will continue sending to the
sub-slave the other half of the command, which is not valid.
The fix, designed by @yangsiran and @antirez, and implemented by
@antirez, uses a secondary buffer in order to feed the sub-masters and
update the replication backlog and offsets, only when a given part of
the query buffer is actually *applied* to the state of the instance,
that is, when the command gets processed and the command is not pending
in the Redis transaction buffer because of CLIENT_MULTI state.
Given that now the backlog and offsets representation are in agreement
with the actual processed commands, both issue 1 and 2 should no longer
be possible.
Thanks to @kevinmcgehee, @yangsiran and @oranagra for their work in
identifying and designing a fix for this problem.
If a thread unblocks a client blocked in a module command, by using the
RedisMdoule_UnblockClient() API, the event loop may not be awaken until
the next timeout of the multiplexing API or the next unrelated I/O
operation on other clients. We actually want the client to be served
ASAP, so a mechanism is needed in order for the unblocking API to inform
Redis that there is a client to serve ASAP.
This commit fixes the issue using the old trick of the pipe: when a
client needs to be unblocked, a byte is written in a pipe. When we run
the list of clients blocked in modules, we consume all the bytes
written in the pipe. Writes and reads are performed inside the context
of the mutex, so no race is possible in which we consume the bytes that
are actually related to an awake request for a client that should still
be put into the list of clients to unblock.
It was verified that after the fix the server handles the blocked
clients with the expected short delay.
Thanks to @dvirsky for understanding there was such a problem and
reporting it.
This change attempts to switch to an hash function which mitigates
the effects of the HashDoS attack (denial of service attack trying
to force data structures to worst case behavior) while at the same time
providing Redis with an hash function that does not expect the input
data to be word aligned, a condition no longer true now that sds.c
strings have a varialbe length header.
Note that it is possible sometimes that even using an hash function
for which collisions cannot be generated without knowing the seed,
special implementation details or the exposure of the seed in an
indirect way (for example the ability to add elements to a Set and
check the return in which Redis returns them with SMEMBERS) may
make the attacker's life simpler in the process of trying to guess
the correct seed, however the next step would be to switch to a
log(N) data structure when too many items in a single bucket are
detected: this seems like an overkill in the case of Redis.
SPEED REGRESION TESTS:
In order to verify that switching from MurmurHash to SipHash had
no impact on speed, a set of benchmarks involving fast insertion
of 5 million of keys were performed.
The result shows Redis with SipHash in high pipelining conditions
to be about 4% slower compared to using the previous hash function.
However this could partially be related to the fact that the current
implementation does not attempt to hash whole words at a time but
reads single bytes, in order to have an output which is endian-netural
and at the same time working on systems where unaligned memory accesses
are a problem.
Further X86 specific optimizations should be tested, the function
may easily get at the same level of MurMurHash2 if a few optimizations
are performed.
This is of great interest because allows us to print debugging
informations that could be of useful when debugging, like in the
following example:
serverPanic("Unexpected encoding for object %d, %d",
obj->type, obj->encoding);
You can still force the logo in the normal logs.
For motivations, check issue #3112. For me the reason is that actually
the logo is nice to have in interactive sessions, but inside the logs
kinda loses its usefulness, but for the ability of users to recognize
restarts easily: for this reason the new startup sequence shows a one
liner ASCII "wave" so that there is still a bit of visual clue.
Startup logging was modified in order to log events in more obvious
ways, and to log more events. Also certain important informations are
now more easy to parse/grep since they are printed in field=value style.
The option --always-show-logo in redis.conf was added, defaulting to no.
The new algorithm provides the same speed with a smaller error for
cardinalities in the range 0-100k. Before switching, the new and old
algorithm behavior was studied in details in the context of
issue #3677. You can find a few graphs and motivations there.
Salvatore Sanfilippo