Further details from @antirez:
It was reported by @StopForumSpam on Twitter that the Redis replication
link was strangely using multiple TCP packets for multiple commands.
This wastes a lot of bandwidth and is due to the TCP_NODELAY option we
enable on the socket after accepting a new connection.
However the master -> slave channel is a one-way channel since Redis
replication is asynchronous, so there is no point in trying to reduce
the latency, we should aim to reduce the bandwidth. For this reason this
commit introduces the ability to disable the nagle algorithm on the
socket after a successful SYNC.
This feature is off by default because the delay can be up to 40
milliseconds with normally configured Linux kernels.
When keyspace events are enabled, the overhead is not sever but
noticeable, so this commit introduces the ability to select subclasses
of events in order to avoid to generate events the user is not
interested in.
The events can be selected using redis.conf or CONFIG SET / GET.
decrRefCount used to get its argument as a void* pointer in order to be
used as destructor where a 'void free_object(void*)' prototype is
expected. However this made simpler to introduce bugs by freeing the
wrong pointer. This commit fixes the argument type and introduces a new
wrapper called decrRefCountVoid() that can be used when the void*
argument is needed.
Sometimes it is much simpler to debug complex Redis installations if it
is possible to assign clients a name that is displayed in the CLIENT
LIST output.
This is the case, for example, for "leaked" connections. The ability to
provide a name to the client makes it quite trivial to understand what
is the part of the code implementing the client not releasing the
resources appropriately.
Behavior:
CLIENT SETNAME: set a name for the client, or remove the current
name if an empty name is set.
CLIENT GETNAME: get the current name, or a nil.
CLIENT LIST: now displays the client name if any.
Thanks to Mark Gravell for pushing this idea forward.
REDIS_HZ is the frequency our serverCron() function is called with.
A more frequent call to this function results into less latency when the
server is trying to handle very expansive background operations like
mass expires of a lot of keys at the same time.
Redis 2.4 used to have an HZ of 10. This was good enough with almost
every setup, but the incremental key expiration algorithm was working a
bit better under *extreme* pressure when HZ was set to 100 for Redis
2.6.
However for most users a latency spike of 30 milliseconds when million
of keys are expiring at the same time is acceptable, on the other hand a
default HZ of 100 in Redis 2.6 was causing idle instances to use some
CPU time compared to Redis 2.4. The CPU usage was in the order of 0.3%
for an idle instance, however this is a shame as more energy is consumed
by the server, if not important resources.
This commit introduces HZ as a runtime parameter, that can be queried by
INFO or CONFIG GET, and can be modified with CONFIG SET. At the same
time the default frequency is set back to 10.
In this way we default to a sane value of 10, but allows users to
easily switch to values up to 500 for near real-time applications if
needed and if they are willing to pay this small CPU usage penalty.
To store the keys we block for during a blocking pop operation, in the
case the client is blocked for more data to arrive, we used a simple
linear array of redis objects, in the blockingState structure:
robj **keys;
int count;
However in order to fix issue #801 we also use a dictionary in order to
avoid to end in the blocked clients queue for the same key multiple
times with the same client.
The dictionary was only temporary, just to avoid duplicates, but since
we create / destroy it there is no point in doing this duplicated work,
so this commit simply use a dictionary as the main structure to store
the keys we are blocked for. So instead of the previous fields we now
just have:
dict *keys;
This simplifies the code and reduces the work done by the server during
a blocking POP operation.
The idea is to be able to identify a build in a unique way, so for
instance after a bug report we can recognize that the build is the one
of a popular Linux distribution and perform the debugging in the same
environment.
EVALSHA used to crash if the SHA1 was not lowercase (Issue #783).
Fixed using a case insensitive dictionary type for the sha -> script
map used for replication of scripts.
After the transcation starts with a MULIT, the previous behavior was to
return an error on problems such as maxmemory limit reached. But still
to execute the transaction with the subset of queued commands on EXEC.
While it is true that the client was able to check for errors
distinguish QUEUED by an error reply, MULTI/EXEC in most client
implementations uses pipelining for speed, so all the commands and EXEC
are sent without caring about replies.
With this change:
1) EXEC fails if at least one command was not queued because of an
error. The EXECABORT error is used.
2) A generic error is always reported on EXEC.
3) The client DISCARDs the MULTI state after a failed EXEC, otherwise
pipelining multiple transactions would be basically impossible:
After a failed EXEC the next transaction would be simply queued as
the tail of the previous transaction.
By caching TCP connections used by MIGRATE to chat with other Redis
instances a 5x performance improvement was measured with
redis-benchmark against small keys.
This can dramatically speedup cluster resharding and other processes
where an high load of MIGRATE commands are used.
Before of this commit it used to be like this:
MULTI
EXEC
... actual commands of the transaction ...
Because after all that is the natural order of things. Transaction
commands are queued and executed *only after* EXEC is called.
However this makes debugging with MONITOR a mess, so the code was
modified to provide a coherent output.
What happens is that MULTI is rendered in the MONITOR output as far as
possible, instead EXEC is propagated only after the transaction is
executed, or even in the case it fails because of WATCH, so in this case
you'll simply see:
MULTI
EXEC
An empty transaction.
Redis provides support for blocking operations such as BLPOP or BRPOP.
This operations are identical to normal LPOP and RPOP operations as long
as there are elements in the target list, but if the list is empty they
block waiting for new data to arrive to the list.
All the clients blocked waiting for th same list are served in a FIFO
way, so the first that blocked is the first to be served when there is
more data pushed by another client into the list.
The previous implementation of blocking operations was conceived to
serve clients in the context of push operations. For for instance:
1) There is a client "A" blocked on list "foo".
2) The client "B" performs `LPUSH foo somevalue`.
3) The client "A" is served in the context of the "B" LPUSH,
synchronously.
Processing things in a synchronous way was useful as if "A" pushes a
value that is served by "B", from the point of view of the database is a
NOP (no operation) thing, that is, nothing is replicated, nothing is
written in the AOF file, and so forth.
However later we implemented two things:
1) Variadic LPUSH that could add multiple values to a list in the
context of a single call.
2) BRPOPLPUSH that was a version of BRPOP that also provided a "PUSH"
side effect when receiving data.
This forced us to make the synchronous implementation more complex. If
client "B" is waiting for data, and "A" pushes three elemnents in a
single call, we needed to propagate an LPUSH with a missing argument
in the AOF and replication link. We also needed to make sure to
replicate the LPUSH side of BRPOPLPUSH, but only if in turn did not
happened to serve another blocking client into another list ;)
This were complex but with a few of mutually recursive functions
everything worked as expected... until one day we introduced scripting
in Redis.
Scripting + synchronous blocking operations = Issue #614.
Basically you can't "rewrite" a script to have just a partial effect on
the replicas and AOF file if the script happened to serve a few blocked
clients.
The solution to all this problems, implemented by this commit, is to
change the way we serve blocked clients. Instead of serving the blocked
clients synchronously, in the context of the command performing the PUSH
operation, it is now an asynchronous and iterative process:
1) If a key that has clients blocked waiting for data is the subject of
a list push operation, We simply mark keys as "ready" and put it into a
queue.
2) Every command pushing stuff on lists, as a variadic LPUSH, a script,
or whatever it is, is replicated verbatim without any rewriting.
3) Every time a Redis command, a MULTI/EXEC block, or a script,
completed its execution, we run the list of keys ready to serve blocked
clients (as more data arrived), and process this list serving the
blocked clients.
4) As a result of "3" maybe more keys are ready again for other clients
(as a result of BRPOPLPUSH we may have push operations), so we iterate
back to step "3" if it's needed.
The new code has a much simpler semantics, and a simpler to understand
implementation, with the disadvantage of not being able to "optmize out"
a PUSH+BPOP as a No OP.
This commit will be tested with care before the final merge, more tests
will be added likely.
SORT is able to return (faster than when ordering) unordered output if
the "BY" clause is used with a constant value. However we try to play
well with scripting requirements of determinism providing always sorted
outputs when SORT (and other similar commands) are called by Lua
scripts.
However we used the general mechanism in place in scripting in order to
reorder SORT output, that is, if the command has the "S" flag set, the
Lua scripting engine will take an additional step when converting a
multi bulk reply to Lua value, calling a Lua sorting function.
This is suboptimal as we can do it faster inside SORT itself.
This is also broken as issue #545 shows us: basically when SORT is used
with a constant BY, and additionally also GET is used, the Lua scripting
engine was trying to order the output as a flat array, while it was
actually a list of key-value pairs.
What we do know is to recognized if the caller of SORT is the Lua client
(since we can check this using the REDIS_LUA_CLIENT flag). If so, and if
a "don't sort" condition is triggered by the BY option with a constant
string, we force the lexicographical sorting.
This commit fixes this bug and improves the performance, and at the same
time simplifies the implementation. This does not mean I'm smart today,
it means I was stupid when I committed the original implementation ;)
During the first synchronization step of the replication process, a Redis
slave connects with the master in a non blocking way. However once the
connection is established the replication continues sending the REPLCONF
command, and sometimes the AUTH command if needed. Those commands are
send in a partially blocking way (blocking with timeout in the order of
seconds).
Because it is common for a blocked master to accept connections even if
it is actually not able to reply to the slave requests, it was easy for
a slave to block if the master had serious issues, but was still able to
accept connections in the listening socket.
For this reason we now send an asynchronous PING request just after the
non blocking connection ended in a successful way, and wait for the
reply before to continue with the replication process. It is very
unlikely that a master replying to PING can't reply to the other
commands.
This solution was proposed by Didier Spezia (Thanks!) so that we don't
need to turn all the replication process into a non blocking affair, but
still the probability of a slave blocked is minimal even in the event of
a failing master.
Also we now use getsockopt(SO_ERROR) in order to check errors ASAP
in the event handler, instead of waiting for actual I/O to return an
error.
This commit fixes issue #632.
A Redis slave can now be configured with a priority, that is an integer
number that is shown in INFO output and can be get and set using the
redis.conf file or the CONFIG GET/SET command.
This field is used by Sentinel during slave election. A slave with lower
priority is preferred. A slave with priority zero is never elected (and
is considered to be impossible to elect even if it is the only slave
available).
A next commit will add support in the Sentinel side as well.
This fixes issue #539.
Basically if there is enough free memory the OS may buffer the RDB file
that the slave transfers on disk from the master. The file may
actually be flused on disk at once by the operating system when it gets
closed by Redis, causing the close system call to block for a long time.
This patch is a modified version of one provided by yoav-steinberg of
@garantiadata (the original version was posted in the issue #539
comments), and tries to flush the OS buffers incrementally (every 8 MB
of loaded data).
This commit implements the first, beta quality implementation of Redis
Sentinel, a distributed monitoring system for Redis with notification
and automatic failover capabilities.
More info at http://redis.io/topics/sentinel
Redis loading data from disk, and a Redis slave disconnected from its
master with serve-stale-data disabled, are two conditions where
commands are normally refused by Redis, returning an error.
However there is no reason to disable Pub/Sub commands as well, given
that this layer does not interact with the dataset. To allow Pub/Sub in
as many contexts as possible is especially interesting now that Redis
Sentinel uses Pub/Sub of a Redis master as a communication channel
between Sentinels.
This commit allows Pub/Sub to be used in the above two contexts where
it was previously denied.
Behaves like rdb_last_bgsave_status -- even down to reporting 'ok' when
no rewrite has been done yet. (You might want to check that
aof_last_rewrite_time_sec is not -1.)
REDIS_REPL_PING_SLAVE_PERIOD controls how often the master should
transmit a heartbeat (PING) to its slaves. This period, which defaults
to 10, is measured in seconds.
Redis 2.4 masters used to ping their slaves every ten seconds, just like
it says on the tin.
The Redis 2.6 masters I have been experimenting with, on the other hand,
ping their slaves *every second*. (master_last_io_seconds_ago never
approaches 10.) I think the ping period was inadvertently slashed to
one-tenth of its nominal value around the time REDIS_HZ was introduced.
This commit reintroduces correct ping schedule behaviour.
The REPLCONF command is an internal command (not designed to be directly
used by normal clients) that allows a slave to set some replication
related state in the master before issuing SYNC to start the
replication.
The initial motivation for this command, and the only reason currently
it is used by the implementation, is to let the slave instance
communicate its listening port to the slave, so that the master can
show all the slaves with their listening ports in the "replication"
section of the INFO output.
This allows clients to auto discover and query all the slaves attached
into a master.
Currently only a single option of the REPLCONF command is supported, and
it is called "listening-port", so the slave now starts the replication
process with something like the following chat:
REPLCONF listening-prot 6380
SYNC
Note that this works even if the master is an older version of Redis and
does not understand REPLCONF, because the slave ignores the REPLCONF
error.
In the future REPLCONF can be used for partial replication and other
replication related features where there is the need to exchange
information between master and slave.
NOTE: This commit also fixes a bug: the INFO outout already carried
information about slaves, but the port was broken, and was obtained
with getpeername(2), so it was actually just the ephemeral port used
by the slave to connect to the master as a client.
The way we compared the authentication password using strcmp() allowed
an attacker to gain information about the password using a well known
class of attacks called "timing attacks".
The bug appears to be practically not exploitable in most modern systems
running Redis since even using multiple bytes of differences in the
input at a time instead of one the difference in running time in in the
order of 10 nanoseconds, making it hard to exploit even on LAN. However
attacks always get better so we are providing a fix ASAP.
The new implementation uses two fixed length buffers and a constant time
comparison function, with the goal of:
1) Completely avoid leaking information about the content of the
password, since the comparison is always performed between 512
characters and without conditionals.
2) Partially avoid leaking information about the length of the
password.
About "2" we still have a stage in the code where the real password and
the user provided password are copied in the static buffers, we also run
two strlen() operations against the two inputs, so the running time
of the comparison is a fixed amount plus a time proportional to
LENGTH(A)+LENGTH(B). This means that the absolute time of the operation
performed is still related to the length of the password in some way,
but there is no way to change the input in order to get a difference in
the execution time in the comparison that is not just proportional to
the string provided by the user (because the password length is fixed).
Thus in practical terms the user should try to discover LENGTH(PASSWORD)
looking at the whole execution time of the AUTH command and trying to
guess a proportionality between the whole execution time and the
password length: this appears to be mostly unfeasible in the real world.
Also protecting from this attack is not very useful in the case of Redis
as a brute force attack is anyway feasible if the password is too short,
while with a long password makes it not an issue that the attacker knows
the length.
The ziplist -> hashtable conversion code is triggered every time an hash
value must be promoted to a full hash table because the number or size of
elements reached the threshold.
If a problem in the ziplist causes the same field to be present
multiple times, the assertion of successful addition of the element
inside the hash table will fail, crashing server with a failed
assertion, but providing little information about the problem.
This code adds a new logging function to perform the hex dump of binary
data, and makes sure that the ziplist -> hashtable conversion code uses
this new logging facility to dump the content of the ziplist when the
assertion fails.
This change was originally made in order to investigate issue #547.
The 'persistence' section of INFO output now contains additional four
fields related to RDB and AOF persistence:
rdb_last_bgsave_time_sec Duration of latest BGSAVE in sec.
rdb_current_bgsave_time_sec Duration of current BGSAVE in sec.
aof_last_rewrite_time_sec Duration of latest AOF rewrite in sec.
aof_current_rewrite_time_sec Duration of current AOF rewrite in sec.
The 'current' fields are set to -1 if a BGSAVE / AOF rewrite is not in
progress. The 'last' fileds are set to -1 if no previous BGSAVE / AOF
rewrites were performed.
Additionally a few fields in the persistence section were renamed for
consistency:
changes_since_last_save -> rdb_changes_since_last_save
bgsave_in_progress -> rdb_bgsave_in_progress
last_save_time -> rdb_last_save_time
last_bgsave_status -> rdb_last_bgsave_status
bgrewriteaof_in_progress -> aof_rewrite_in_progress
bgrewriteaof_scheduled -> aof_rewrite_scheduled
After the renaming, fields in the persistence section start with rdb_ or
aof_ prefix depending on the persistence method they describe.
The field 'loading' and related fields are not prefixed because they are
unique for both the persistence methods.
The motivation for this new commands is to be search in the usage of
Redis for real time statistics. See the article "Fast real time metrics
using Redis".
http://blog.getspool.com/2011/11/29/fast-easy-realtime-metrics-using-redis-bitmaps/
In general Redis strings when used as bitmaps using the SETBIT/GETBIT
command provide a very space-efficient and fast way to store statistics.
For instance in a web application with users, every user can be
associated with a key that shows every day in which the user visited the
web service. This information can be really valuable to extract user
behaviour information.
With Redis bitmaps doing this is very simple just saying that a given
day is 0 (the data the service was put online) and all the next days are
1, 2, 3, and so forth. So with SETBIT it is possible to set the bit
corresponding to the current day every time the user visits the site.
It is possible to take the count of the bit sets on the run, this is
extremely easy using a Lua script. However a fast bit count native
operation can be useful, especially if it can operate on ranges, or when
the string is small like in the case of days (even if you consider many
years it is still extremely little data).
For this reason BITOP was introduced. The command counts the number of
bits set to 1 in a string, with optional range:
BITCOUNT key [start end]
The start/end parameters are similar to GETRANGE. If omitted the whole
string is tested.
Population counting is more useful when bit-level operations like AND,
OR and XOR are avaialble. For instance I can test multiple users to see
the number of days three users visited the site at the same time. To do
this we can take the AND of all the bitmaps, and then count the set bits.
For this reason the BITOP command was introduced:
BITOP [AND|OR|XOR|NOT] dest_key src_key1 src_key2 src_key3 ... src_keyN
In the special case of NOT (that inverts the bits) only one source key
can be passed.
The judicious use of BITCOUNT and BITOP combined can lead to interesting
use cases with very space efficient representation of data.
The implementation provided is still not tested and optimized for speed,
next commits will introduce unit tests. Later the implementation will be
profiled to see if it is possible to gain an important amount of speed
without making the code much more complex.
During the AOF rewrite process, the parent process needs to accumulate
the new writes in an in-memory buffer: when the child will terminate the
AOF rewriting process this buffer (that ist the difference between the
dataset when the rewrite was started, and the current dataset) is
flushed to the new AOF file.
We used to implement this buffer using an sds.c string, but sds.c has a
2GB limit. Sometimes the dataset can be big enough, the amount of writes
so high, and the rewrite process slow enough that we overflow the 2GB
limit, causing a crash, documented on github by issue #504.
In order to prevent this from happening, this commit introduces a new
system to accumulate writes, implemented by a linked list of blocks of
10 MB each, so that we also avoid paying the reallocation cost.
Note that theoretically modern operating systems may implement realloc()
simply as a remaping of the old pages, thus with very good performances,
see for instance the mremap() syscall on Linux. However this is not
always true, and jemalloc by default avoids doing this because there are
issues with the current implementation of mremap().
For this reason we are using a linked list of blocks instead of a single
block that gets reallocated again and again.
The changes in this commit lacks testing, that will be performed before
merging into the unstable branch. This fix will not enter 2.4 because it
is too invasive. However 2.4 will log a warning when the AOF rewrite
buffer is near to the 2GB limit.
A previous commit introduced REDIS_HZ define that changes the frequency
of calls to the serverCron() Redis function. This commit improves
different related things:
1) Software watchdog: now the minimal period can be set according to
REDIS_HZ. The minimal period is two times the timer period, that is:
(1000/REDIS_HZ)*2 milliseconds
2) The incremental rehashing is now performed in the expires dictionary
as well.
3) The activeExpireCycle() function was improved in different ways:
- Now it checks if it already used too much time using microseconds
instead of milliseconds for better precision.
- The time limit is now calculated correctly, in the previous version
the division was performed before of the multiplication resulting in
a timelimit of 0 if HZ was big enough.
- Databases with less than 1% of buckets fill in the hash table are
skipped, because getting random keys is too expensive in this
condition.
4) tryResizeHashTables() is now called at every timer call, we need to
match the number of calls we do to the expired keys colleciton cycle.
5) REDIS_HZ was raised to 100.
Redis uses a function called serverCron() that is very similar to the
timer interrupt of an operating system. This function is used to handle
a number of asynchronous things, like active expired keys collection,
clients timeouts, update of statistics, things related to the cluster
and replication, triggering of BGSAVE and AOF rewrite process, and so
forth.
In the past the timer was called 1 time per second. At some point it was
raised to 10 times per second, but it still was fixed and could not be
changed even at compile time, because different functions called from
serverCron() assumed a given fixed frequency.
This commmit makes the frequency configurable, so that it is simpler to
pick a good tradeoff between overhead of this function (that is usually
very small) and the responsiveness of Redis during a few critical
circumstances where a lot of work is done inside the timer.
An example of such a critical condition is mass-expire of a lot of keys
in the same second. Up to a given percentage of CPU time is used to
perform expired keys collection per expire cylce. Now changing the
REDIS_HZ macro it is possible to do less work but more times per second
in order to block the server for less time.
If this patch will work well in our tests it will enter Redis 2.6-final.
If a large amonut of keys are all expiring about at the same time, the
"active" expired keys collection cycle used to block as far as the
percentage of already expired keys was >= 25% of the total population of
keys with an expire set.
This could block the server even for many seconds in order to reclaim
memory ASAP. The new algorithm uses at max a small amount of
milliseconds per cycle, even if this means reclaiming the memory less
promptly it also means a more responsive server.
We used to reply -ERR ... message ..., now the reply is
instead -MASTERDOWN ... message ... so that it can be distinguished
easily by the other error conditions.
Two limits are added:
1) Up to SLOWLOG_ENTRY_MAX_ARGV arguments are logged.
2) Up to SLOWLOG_ENTRY_MAX_STRING bytes per argument are logged.
3) slowlog-max-len is set to 128 by default (was 1024).
The number of remaining arguments / bytes is logged in the entry
so that the user can understand better the nature of the logged command.
