This commit attempts to fix a bug involving PSYNC and diskless
replication (currently experimental) found by Yuval Inbar from Redis Labs
and that was later found to have even more far reaching effects (the bug also
exists when diskstore is off).
The gist of the bug is that, a Redis master replies with +FULLRESYNC to
a PSYNC attempt that fails and requires a full resynchronization.
However, the baseline offset sent along with FULLRESYNC was always the
current master replication offset. This is not ok, because there are
many reasosn that may delay the RDB file creation. And... guess what,
the master offset we communicate must be the one of the time the RDB
was created. So for example:
1) When the BGSAVE for replication is delayed since there is one
already but is not good for replication.
2) When the BGSAVE is not needed as we attach one currently ongoing.
3) When because of diskless replication the BGSAVE is delayed.
In all the above cases the PSYNC reply is wrong and the slave may
reconnect later claiming to need a wrong offset: this may cause
data curruption later.
Using chained replication where C is slave of B which is in turn slave of
A, if B reconnects the replication link with A but discovers it is no
longer possible to PSYNC, slaves of B must be disconnected and PSYNC
not allowed, since the new B dataset may be completely different after
the synchronization with the master.
Note that there are varius semantical differences in the way this is
handled now compared to the past. In the past the semantics was:
1. When a slave lost connection with its master, disconnected the chained
slaves ASAP. Which is not needed since after a successful PSYNC with the
master, the slaves can continue and don't need to resync in turn.
2. However after a failed PSYNC the replication backlog was not reset, so a
slave was able to PSYNC successfully even if the instance did a full
sync with its master, containing now an entirely different data set.
Now instead chained slaves are not disconnected when the slave lose the
connection with its master, but only when it is forced to full SYNC with
its master. This means that if the slave having chained slaves does a
successful PSYNC all its slaves can continue without troubles.
See issue #2694 for more details.
We usually want to reach the master using the address of the interface
Redis is bound to (via the "bind" config option). That's useful since
the master will get (and publish) the slave address getting the peer
name of the incoming socket connection from the slave.
However, when this is not possible, for example because the slave is
bound to the loopback interface but repliaces from a master accessed via
an external interface, we want to still connect with the master even
from a different interface: in this case it is not really important that
the master will provide any other address, while it is vital to be able
to replicate correctly.
Related to issues #2609 and #2612.
When we fail to setup the write handler it does not make sense to take
the client around, it is missing writes: whatever is a client or a slave
anyway the connection should terminated ASAP.
Moreover what the function does exactly with its return value, and in
which case the write handler is installed on the socket, was not clear,
so the functions comment are improved to make the goals of the function
more obvious.
Also related to #2485.
master was closing the connection if the RDB transfer took long time.
and also sent PINGs to the slave before it got the initial ACK, in which case the slave wouldn't be able to find the EOF marker.
Bug as old as Redis and blocking operations. It's hard to trigger since
only happens on instance role switch, but the results are quite bad
since an inconsistency between master and slave is created.
How to trigger the bug is a good description of the bug itself.
1. Client does "BLPOP mylist 0" in master.
2. Master is turned into slave, that replicates from New-Master.
3. Client does "LPUSH mylist foo" in New-Master.
4. New-Master propagates write to slave.
5. Slave receives the LPUSH, the blocked client get served.
Now Master "mylist" key has "foo", Slave "mylist" key is empty.
Highlights:
* At step "2" above, the client remains attached, basically escaping any
check performed during command dispatch: read only slave, in that case.
* At step "5" the slave (that was the master), serves the blocked client
consuming a list element, which is not consumed on the master side.
This scenario is technically likely to happen during failovers, however
since Redis Sentinel already disconnects clients using the CLIENT
command when changing the role of the instance, the bug is avoided in
Sentinel deployments.
Closes#2473.
Track bandwidth used by clients and replication (but diskless
replication is not tracked since the actual transfer happens in the
child process).
This includes a refactoring that makes tracking new instantaneous
metrics simpler.
RDB EOF detection was relying on the final part of the RDB transfer to
be a magic 40 bytes EOF marker. However as the slave is put online
immediately, and because of sockets timeouts, the replication stream is
actually contiguous with the RDB file.
This means that to detect the EOF correctly we should either:
1) Scan all the stream searching for the mark. Sucks CPU-wise.
2) Start to send the replication stream only after an acknowledge.
3) Implement a proper chunked encoding.
For now solution "2" was picked, so the master does not start to send
ASAP the stream of commands in the case of diskless replication. We wait
for the first REPLCONF ACK command from the slave, that certifies us
that the slave correctly loaded the RDB file and is ready to get more
data.
Same as the original bind fixes (we just missed these the
first time around).
This helps Redis not automatically send
connections from the first IP on an interface if we are bound
to a specific IP address (e.g. with multiple IP aliases on one
interface, you want to send from _your_ IP, not from the first IP
on the interface).
This caused BGSAVE to be triggered a second time without any need when
we switch from socket to disk target via the command
CONFIG SET repl-diskless-sync no
and there is already a slave waiting for the BGSAVE to start.
Also comments clarified about what is happening.
This is useful for normal replication in order to refresh the slave
when we are persisting on disk, but for diskless replication the
child is already receiving data while in WAIT_BGSAVE_END state.
If we turn from diskless to disk-based replication via CONFIG SET, we
need a way to start a BGSAVE if there are slaves alerady waiting for a
BGSAVE to start. Normally with disk-based replication we do it as soon
as the previous child exits, but when there is a configuration change
via CONFIG SET, we may have slaves in WAIT_BGSAVE_START state without
an RDB background process currently active.
The new ROLE command is designed in order to provide a client with
informations about the replication in a fast and easy to use way
compared to the INFO command where the same information is also
available.
This commit adds peer ID caching in the client structure plus an API
change and the use of sdsMakeRoomFor() in order to improve the
reallocation pattern to generate the CLIENT LIST output.
Both the changes account for a very significant speedup.
Sometime an osx master with a Linux server over a slow link caused
a strange error where osx called the writable function for
the socket but actually apparently there was no room in the socket
buffer to accept the write: write(2) call returned an EAGAIN error,
that was not checked, so we considered write(2) == 0 always as a connection
reset, which was unfortunate since the bulk transfer has to start again.
Also more errors are logged with the WARNING level in the same code path
now.
Return the number of slaves for the same master having a better
replication offset of the current slave, that is, the slave "rank" used
to pick a delay before the request for election.
When an instance is potentially set to replicate with another master, it
is conceptually disconnected forever, since we have no old copy of the
dataset for this master in memory.
Masters not understanding REPLCONF ACK will reply with errors to our
requests causing a number of possible issues.
This commit detects a global replication offest set to -1 at the end of
the replication, and marks the client representing the master with the
REDIS_PRE_PSYNC flag.
Note that this flag was called REDIS_PRE_PSYNC_SLAVE but now it is just
REDIS_PRE_PSYNC as it is used for both slaves and masters starting with
this commit.
This commit fixes issue #1488.
Currently replication offsets could be used into a limited way in order
to understand, out of a set of slaves, what is the one with the most
updated data. For example this comparison is possible of N slaves
were replicating all with the same master.
However the replication offset was not transferred from master to slaves
(that are later promoted as masters) in any way, so for instance if
there were three instances A, B, C, with A master and B and C
replication from A, the following could happen:
C disconnects from A.
B is turned into master.
A is switched to master of B.
B receives some write.
In this context there was no way to compare the offset of A and C,
because B would use its own local master replication offset as
replication offset to initialize the replication with A.
With this commit what happens is that when B is turned into master it
inherits the replication offset from A, making A and C comparable.
In the above case assuming no inconsistencies are created during the
disconnection and failover process, A will show to have a replication
offset greater than C.
Note that this does not mean offsets are always comparable to understand
what is, in a set of instances, since in more complex examples the
replica with the higher replication offset could be partitioned away
when picking the instance to elect as new master. However this in
general improves the ability of a system to try to pick a good replica
to promote to master.
The previous fix for false positive timeout detected by master was not
complete. There is another blocking stage while loading data for the
first synchronization with the master, that is, flushing away the
current data from the DB memory.
This commit uses the newly introduced dict.c callback in order to make
some incremental work (to send "\n" heartbeats to the master) while
flushing the old data from memory.
It is hard to write a regression test for this issue unfortunately. More
support for debugging in the Redis core would be needed in terms of
functionalities to simulate a slow DB loading / deletion.
Redis hash table implementation has many non-blocking features like
incremental rehashing, however while deleting a large hash table there
was no way to have a callback called to do some incremental work.
This commit adds this support, as an optiona callback argument to
dictEmpty() that is currently called at a fixed interval (one time every
65k deletions).
Sometimes when we resurrect a cached master after a successful partial
resynchronization attempt, there is pending data in the output buffers
of the client structure representing the master (likely REPLCONF ACK
commands).
If we don't reinstall the write handler, it will never be installed
again by addReply*() family functions as they'll assume that if there is
already data pending, the write handler is already installed.
This bug caused some slaves after a successful partial sync to never
send REPLCONF ACK, and continuously being detected as timing out by the
master, with a disconnection / reconnection loop.
There was a bug that over-esteemed the amount of backlog available,
however this could only happen when a slave was asking for an offset
that was in the "future" compared to the master replication backlog.
Now this case is handled well and logged as an incident in the master
log file.
The previous code using a static buffer as an optimization was lame:
1) Premature optimization, actually it was *slower* than naive code
because resulted into the creation / destruction of the object
encapsulating the output buffer.
2) The code was very hard to test, since it was needed to have specific
tests for command lines exceeding the size of the static buffer.
3) As a result of "2" the code was bugged as the current tests were not
able to stress specific corner cases.
It was replaced with easy to understand code that is safer and faster.
During the replication full resynchronization process, the RDB file is
transfered from the master to the slave. However there is a short
preamble to send, that is currently just the bulk payload length of the
file in the usual Redis form $..length..<CR><LF>.
This preamble used to be sent with a direct write call, assuming that
there was alway room in the socket output buffer to hold the few bytes
needed, however this does not scale in case we'll need to send more
stuff, and is not very robust code in general.
This commit introduces a more general mechanism to send a preamble up to
2GB in size (the max length of an sds string) in a non blocking way.
Clients using SYNC to replicate are older implementations, such as
redis-cli --slave, and are not designed to acknowledge the master with
REPLCONF ACK commands, so we don't have any feedback and should not
disconnect them on timeout.
This code is only responsible to take an LRU-evicted fixed length cache
of SHA1 that we are sure all the slaves received.
In this commit only the implementation is provided, but the Redis core
does not use it to actually send EVALSHA to slaves when possible.
This feature allows the user to specify the minimum number of
connected replicas having a lag less or equal than the specified
amount of seconds for writes to be accepted.
Now masters, using the time at which the last REPLCONF ACK was received,
are able to explicitly disconnect slaves that are no longer responding.
Previously the only chance was to see a very long output buffer, that
was highly suboptimal.
ACKs can be also used as a base for synchronous replication. However in
that case they'll be explicitly requested by the master when the client
sends a request that needs to be replicated synchronously.
This special command is used by the slave to inform the master the
amount of replication stream it currently consumed.
it does not return anything so that we not need to consume additional
bandwidth needed by the master to reply something.
The master can do a number of things knowing the amount of stream
processed, such as understanding the "lag" in bytes of the slave, verify
if a given command was already processed by the slave, and so forth.
When we are preparing an handshake with the slave we can't touch the
connection buffer as it'll be used to accumulate differences between
the sent RDB file and what arrives next from clients.
So in short we can't use addReply() family functions.
However we just use write(2) because we know that the socket buffer is
empty, since a prerequisite for SYNC to work is that the static buffer
and the output list are empty, and in general it is not expected that a
client SYNCs after doing some heavy I/O with the master.
However a short write connection is explicitly handled to avoid
fragility (we simply close the connection and the slave will retry).
SELECT was still transmitted to slaves using the inline protocol, that
is conceived mostly for humans to type into telnet sessions, and is
notably not understood by redis-cli --slave.
Now the new protocol is used instead.
A Redis master sends PING commands to slaves from time to time: doing
this ensures that even if absence of writes, the master->slave channel
remains active and the slave can feel the master presence, instead of
closing the connection for timeout.
This commit changes the way PINGs are sent to slaves in order to use the
standard interface used to replicate all the other commands, that is,
the function replicationFeedSlaves().
With this change the stream of commands sent to every slave is exactly
the same regardless of their exact state (Transferring RDB for first
synchronization or slave already online). With the previous
implementation the PING was only sent to online slaves, with the result
that the output stream from master to slaves was not identical for all
the slaves: this is a problem if we want to implement partial resyncs in
the future using a global replication stream offset.
TL;DR: this commit should not change the behaviour in practical terms,
but is just something in preparation for partial resynchronization
support.
Before this commit every Redis slave had its own selected database ID
state. This was not actually useful as the emitted stream of commands
is identical for all the slaves.
Now the the currently selected database is a global state that is set to
-1 when a new slave is attached, in order to force the SELECT command to
be re-emitted for all the slaves.
This change is useful in order to implement replication partial
resynchronization in the future, as makes sure that the stream of
commands received by slaves, including SELECT commands, are exactly the
same for every slave connected, at any time.
In this way we could have a global offset that can identify a specific
piece of the master -> slaves stream of commands.
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.
Issue #828 shows how Redis was not correctly undoing a non-blocking
connection attempt with the previous master when the master was set to a
new address using the SLAVEOF command.
This was also a result of lack of refactoring, so now there is a
function to cancel the non blocking handshake with the master.
The new function is now used when SLAVEOF NO ONE is called or when
SLAVEOF is used to set the master to a different address.
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.
The new message now contains an hint about modifying the repl-timeout
configuration directive if the problem persists.
This should normally not be needed, because while the master generates
the RDB file it makes sure to send newlines to the replication channel
to prevent timeouts. However there are times when masters running on
very slow systems can completely stop for seconds during the RDB saving
process. In such a case enlarging the timeout value can fix the problem.
See issue #695 for an example of this problem in an EC2 deployment.
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.
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).
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 user @jokea noticed that the following line of code into
replication.c made little sense:
addReplySds(slave,sdsempty());
Investigating a bit I found that this was introduced by commit 6208b3a7
three years ago in the early stages of Redis. The code apparently is not
useful at all, so I'm removing it.
This change will not be backported into 2.4 so that in the rare case
this should introduce a bug, we'll have a chance to detect it into the
development branch. However following the code path it seems like the
code is not useful at all, so the risk is truly small.
