As Oran Agra suggested, in startBgsaveForReplication() when the BGSAVE
attempt returns an error, we scan the list of slaves in order to remove
them since there is no way to serve them currently.
However we check for the replication state BGSAVE_START, which was
modified by rdbSaveToSlaveSockets() before forking(). So when fork fails
the state of slaves remain BGSAVE_END and no cleanup is performed.
This commit fixes the problem by making rdbSaveToSlavesSockets() able to
undo the state change on fork failure.
We have a check to rewrite the config properly when a failover is in
progress, in order to add the current (already failed over) master as
slave, and don't include in the slave list the promoted slave itself.
However there was an issue, the variable with the right address was
computed but never used when the code was modified, and no tests are
available for this feature for two reasons:
1. The Sentinel unit test currently does not test Sentinel ability to
persist its state at all.
2. It is a very hard to trigger state since it lasts for little time in
the context of the testing framework.
However this feature should be covered in the test in some way.
The bug was found by @badboy using the clang static analyzer.
Effects of the bug on safety of Sentinel
===
This bug results in severe issues in the following case:
1. A Sentinel is elected leader.
2. During the failover, it persists a wrong config with a known-slave
entry listing the master address.
3. The Sentinel crashes and restarts, reading invalid configuration from
disk.
4. It sees that the slave now does not obey the logical configuration
(should replicate from the current master), so it sends a SLAVEOF
command to the master (since the slave master is the same) creating a
replication loop (attempt to replicate from itself) which Redis is
currently unable to detect.
5. This means that the master is no longer available because of the bug.
However the lack of availability should be only transient (at least
in my tests, but other states could be possible where the problem
is not recovered automatically) because:
6. Sentinels treat masters reporting to be slaves as failing.
7. A new failover is triggered, and a slave is promoted to master.
Bug lifetime
===
The bug is there forever. Commit 16237d78 actually tried to fix the bug
but in the wrong way (the computed variable was never used! My fault).
So this bug is there basically since the start of Sentinel.
Since the bug is hard to trigger, I remember little reports matching
this condition, but I remember at least a few. Also in automated tests
where instances were stopped and restarted multiple times automatically
I remember hitting this issue, however I was not able to reproduce nor
to determine with the information I had at the time what was causing the
issue.
Before this commit, after triggering a BGSAVE it was up to the caller of
startBgsavForReplication() to handle slaves in WAIT_BGSAVE_START in
order to update them accordingly. However when the replication target is
the socket, this is not possible since the process of updating the
slaves and sending the FULLRESYNC reply must be coupled with the process
of starting an RDB save (the reason is, we need to send the FULLSYNC
command and spawn a child that will start to send RDB data to the slaves
ASAP).
This commit moves the responsibility of handling slaves in
WAIT_BGSAVE_START to startBgsavForReplication() so that for both
diskless and disk-based replication we have the same chain of
responsiblity. In order accomodate such change, the syncCommand() also
needs to put the client in the slave list ASAP (just after the initial
checks) and not at the end, so that startBgsavForReplication() can find
the new slave alrady in the list.
Another related change is what happens if the BGSAVE fails because of
fork() or other errors: we now remove the slave from the list of slaves
and send an error, scheduling the slave connection to be terminated.
As a side effect of this change the following errors found by
Oran Agra are fixed (thanks!):
1. rdbSaveToSlavesSockets() on failed fork will get the slaves cleaned
up, otherwise they remain in a wrong state forever since we setup them
for full resync before actually trying to fork.
2. updateSlavesWaitingBgsave() with replication target set as "socket"
was broken since the function changed the slaves state from
WAIT_BGSAVE_START to WAIT_BGSAVE_END via
replicationSetupSlaveForFullResync(), so later rdbSaveToSlavesSockets()
will not find any slave in the right state (WAIT_BGSAVE_START) to feed.
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.
Talking with @oranagra we had to reason a little bit to understand if
this function could ever flush the output buffers of the wrong slaves,
having online state but actually not being ready to receive writes
before the first ACK is received from them (this happens with diskless
replication).
Next time we'll just read this comment.
It is simpler if removing the read event handler from the FD is up to
slaveTryPartialResynchronization, after all it is only called in the
context of syncWithMaster.
This commit also makes sure that on error all the event handlers are
removed from the socket before closing it.
Add the concept of slaves capabilities to Redis, the slave now presents
to the Redis master with a set of capabilities in the form:
REPLCONF capa SOMECAPA capa OTHERCAPA ...
This has the effect of setting slave->slave_capa with the corresponding
SLAVE_CAPA macros that the master can test later to understand if it
the slave will understand certain formats and protocols of the
replication process. This makes it much simpler to introduce new
replication capabilities in the future in a way that don't break old
slaves or masters.
This patch was designed and implemented together with Oran Agra
(@oranagra).
Our function to read a line with a timeout handles newlines as requests
to refresh the timeout, however the code kept subtracting the buffer
size left every time a newline was received, for a bug in the loop
logic. Fixed by this commit.
For PINGs we use the period configured by the user, but for the newlines
of slaves waiting for an RDB to be created (including slaves waiting for
the FULLRESYNC reply) we need to ping with frequency of 1 second, since
the timeout is fixed and needs to be refreshed.
In previous commits we moved the FULLRESYNC to the moment we start the
BGSAVE, so that the offset we provide is the right one. However this
also means that we need to re-emit the SELECT statement every time a new
slave starts to accumulate the changes.
To obtian this effect in a more clean way, the function that sends the
FULLRESYNC reply was overloaded with a more important role of also doing
this and chanigng the slave state. So it was renamed to
replicationSetupSlaveForFullResync() to better reflect what it does now.
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.
The previos attempt to process each client at least once every ten
seconds was not a good idea, because:
1. Usually because of the past min iterations set to 50, you get much
better processing period most of the times.
2. However when there are many clients and a normal setting for
server.hz, the edge case is triggered, and waiting 10 seconds for a
BLPOP that asked for 1 second is not ok.
3. Moreover, because of the high min-itereations limit of 50, when HZ
was set to an high value, the actual behavior was to process a lot of
clients per second.
Also the function checking for timeouts called gettimeofday() at each
iteration which can be costly.
The new implementation will try to process each client once per second,
gets the current time as argument, and does not attempt to process more
than 5 clients per iteration if not needed.
So now:
1. The CPU usage of an idle Redis process is the same or better.
2. The CPU usage of a busy Redis process is the same or better.
3. However a non trivial amount of work may be performed per iteration
when there are many many clients. In this particular case the user may
want to raise the "HZ" value if needed.
Btw with 4000 clients it was still not possible to noticy any actual
latency created by processing 400 clients per second, since the work
performed for each client is pretty small.
The new return value is the number of keys existing, among the ones
specified in the command line, counting the same key multiple times if
given multiple times (and if it exists).
See PR #2667.
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.
This performs a best effort source address binding attempt. If it is
possible to bind the local address and still have a successful
connect(), then this socket is returned. Otherwise the call is retried
without source address binding attempt.
Related to issues #2609 and #2612.
