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.
When empty strings are created, or when sdsMakeRoomFor() is called, we
are likely into an appending pattern. Use at least type 8 SDS strings
since TYPE 5 does not remember the free allocation size and requires to
call sdsMakeRoomFor() at every new piece appended.
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.
This is an attempt to use the refcount feature of the sds.c fork
provided in the Pull Request #2509. A new type, SDS_TYPE_5 is introduced
having a one byte header with just the string length, without
information about the available additional length at the end of the
string (this means that sdsMakeRoomFor() will be required each time
we want to append something, since the string will always report to have
0 bytes available).
More work needed in order to avoid common SDS functions will pay the
cost of this type. For example both sdscatprintf() and sdscatfmt()
should try to upgrade to SDS_TYPE_8 ASAP when appending chars.
The command reports information about the hash table internal state
representing the specified database ID.
This can be used in order to investigate rehashings, memory usage issues
and for other debugging purposes.
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.
Rationale:
1. The commands look like internals exposed without a real strong use
case.
2. Whatever there is an use case, the client would implement the
commands client side instead of paying RTT just to use a simple to
reimplement library.
3. They add complexity to an otherwise quite straightforward API.
So for now KILLED ;-)
Instead of successive divisions in iteration the new code uses bitwise
magic to interleave / deinterleave two 32bit values into a 64bit one.
All tests still passing and is measurably faster, so worth it.
Stack traces produced by Redis on crash are the most useful tool we
have to fix non easily reproducible crashes, or even easily reproducible
ones where the user just posts a bug report and does not collaborate
furhter.
By declaring functions "static" they no longer show up in the stack
trace.
If GEOENCODE must be our door to enter the Geocoding implementation
details and do fancy things client side, than return the scores as well
so that we can query the sorted sets directly if we wish to do the same
search multiple times, or want to compute the boxes in the client side
to refine our search needs.
The GIS standard and all the major DBs implementing GIS related
functions take coordinates as x,y that is longitude,latitude.
It was a bad start for Redis to do things differently, so even if this
means that existing users of the Geo module will be required to change
their code, Redis now conforms to the standard.
Usually Redis is very backward compatible, but this is not an exception
to this rule, since this is the first Geo implementation entering the
official Redis source code. It is not wise to try to be backward
compatible with code forks... :-)
Close#2637.
The returned step was in some case not enough towards normal
coordinates (for example when our search position was was already near the
margin of the central area, and we had to match, using the east or west
neighbor, a very far point). Example:
geoadd points 67.575457940146066 -62.001317572780565 far
geoadd points 66.685439060295664 -58.925040587282297 center
georadius points 66.685439060295664 -58.925040587282297 200 km
In the above case the code failed to find a match (happens at smaller
latitudes too) even if far and center are at less than 200km.
Another fix introduced by this commit is a progressively larger area
towards the poles, since meridians are a lot less far away, so we need
to compensate for this.
The current implementation works comparably to the Tcl brute-force
stress tester implemented in the fuzzy test in the geo.tcl unit for
latitudes between -70 and 70, and is pretty accurate over +/-80 too,
with sporadic false negatives.
A more mathematically clean implementation is possible by computing the
meridian distance at the specified latitude and computing the step
according to it.
1. We no longer use a fake client but just rewriting.
2. We group all the inserts into a single ZADD dispatch (big speed win).
3. As a side effect of the correct implementation, replication works.
4. The return value of the command is now correct.
This commit simplifies the implementation in a few ways:
1. zsetScore implementation improved a bit and moved into t_zset.c where
is now also used to implement the ZSCORE command.
2. Range extraction from the sorted set remains a separated
implementation from the one in t_zset.c, but was hyper-specialized in
order to avoid accumulating results into a list and remove the ones
outside the radius.
3. A new type is introduced: geoArray, which can accumulate geoPoint
structures in a vector with power of two expansion policy. This is
useful since we have to call qsort() against it before returning the
result to the user.
4. As a result of 1, 2, 3, the two files zset.c and zset.h are now
removed, including the function to merge two lists (now handled with
functions that can add elements to existing geoArray arrays) and
the machinery used in order to pass zset results.
5. geoPoint structure simplified because of the general code structure
simplification, so we no longer need to take references to objects.
6. Not counting the JSON removal the refactoring removes 200 lines of
code for the same functionalities, with a simpler to read
implementation.
7. GEORADIUS is now 2.5 times faster testing with 10k elements and a
radius resulting in 124 elements returned. However this is mostly a
side effect of the refactoring and simplification. More speed gains
can be achieved by trying to optimize the code.
For some reason the Geo PR included disabling the fact that Redis is
compiled with optimizations. Apparently it was just @mattsta attempt to
speedup the modify-compile-test iteration and there are no other
reasons.
This feature apparently is not going to be very useful, to send a
GEOADD+PUBLISH combo is exactly the same. One that would make a ton of
difference is the ability to subscribe to a position and a radius, and
get the updates in terms of objects entering/exiting the area.
Current todo:
- replace functions in zset.{c,h} with a new unified Redis
zset access API.
Once we get the zset interface fixed, we can squash
relevant commits in this branch and have one nice commit
to merge into unstable.
This commit adds:
- Geo commands
- Tests; runnable with: ./runtest --single unit/geo
- Geo helpers in deps/geohash-int/
- src/geo.{c,h} and src/geojson.{c,h} implementing geo commands
- Updated build configurations to get everything working
- TEMPORARY: src/zset.{c,h} implementing zset score and zset
range reading without writing to client output buffers.
- Modified linkage of one t_zset.c function for use in zset.c
Conflicts:
src/Makefile
src/redis.c
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.
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.
Normally ZADD only returns the number of elements added to a sorted
set, using the RETCH option it returns the sum of elements added or
for which the score was updated.
The PING trigger was improved again by using two fields instead of a
single one to remember when the last ping was sent:
1. The "active" ping is the time at which we sent the last ping that
still received no reply. However we continue to ping non replying
instances even if they have an old active ping: the link may be
disconnected and reconencted in the meantime so the older pings may get
lost even if it's a TCP socket.
2. The "last" ping is the time at which we really sent the last ping
on the wire, and this is used in order to throttle the amount of pings
we send during failures (when no pong is received).
All in all the failure detector effectiveness should be identical but we
avoid to flood instances with pings during failures or when they are
slow.
It's ok to ping as soon as the ping period has elapsed since we received
the last PONG, but it's not good that we ping again if there is a
pending ping... With this change we'll send a new ping if there is one
pending only if two times the ping period elapsed since the ping which
is still pending was sent.
This is useful for debugging and logging activities: given a
sentinelRedisInstance object returns a C string representing the
instance type: master, slave, sentinel.
This new command triggers a config flush to save the in-memory config to
disk. This is useful for cases of a configuration management system or a
package manager wiping out your sentinel config while the process is
still running - and has not yet been restarted. It can also be useful
for scripting a backup and migrate or clone of a running sentinel.
Since with a previous commit Sentinels now persist their unique ID, we
no longer need to detect duplicated Sentinels and re-add them. We remove
and re-add back using different events only in the case of address
switch of the same Sentinel, without generating a new +sentinel event.
Previously Sentinels always changed unique ID across restarts, relying
on the server.runid field. This is not a good idea, and forced Sentinel
to rely on detection of duplicated Sentinels and a potentially dangerous
clean-up and re-add operation of the Sentinel instance that was
rebooted.
Now the ID is generated at the first start and persisted in the
configuration file, so that a given Sentinel will have its unique
ID forever (unless the configuration is manually deleted or there is a
filesystem corruption).
Originally, only the +slave event which occurs when a slave is
reconfigured during sentinelResetMasterAndChangeAddress triggers a flush
of the config to disk. However, newly discovered slaves don't
apparently trigger this flush but do trigger the +slave event issuance.
So if you start up a sentinel, add a master, then add a slave to the
master (as a way to reproduce it) you'll see the +slave event issued,
but the sentinel config won't be updated with the known-slave entry.
This change makes sentinel do the flush of the config if a new slave is
deteted in sentinelRefreshInstanceInfo.
To rewrite the config in the loop that adds slaves back after a master
reset, in order to handle switching to another master, is useless: it
just adds latency since there is an fsync call in the inner loop,
without providing any additional guarantee, but the contrary, since if
after the first loop iteration the server crashes we end with just a
single slave entry losing all the other informations.
It is wiser to rewrite the config at the end when the full new
state is configured.
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.
Segfault introduced during a refactoring / warning suppression a few
commits away. This particular call assumed that it is safe to pass NULL
to the object pointer argument when we are sure the set has a given
encoding. This can't be assumed and is now guaranteed to segfault
because of the new API of setTypeNext().
This change fixes several warnings compiling at -O3 level with GCC
4.8.2, and at the same time, in case of misuse of the API, we have the
pointer initialize to NULL or the integer initialized to the value
-123456789 which is easy to spot by naked eye.
No semantical changes since to make dict.c truly able to scale over the
32 bit table size limit, the hash function shoulds and other internals
related to hash function output should be 64 bit ready.
rehashidx is always positive in the two code paths, since the only
negative value it could have is -1 when there is no rehashing in
progress, and the condition is explicitly checked.
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.
There was a bug in Redis Cluster caused by clients blocked in a blocking
list pop operation, for keys no longer handled by the instance, or
in a condition where the cluster became down after the client blocked.
A typical situation is:
1) BLPOP <somekey> 0
2) <somekey> hash slot is resharded to another master.
The client will block forever int this case.
A symmentrical non-cluster-specific bug happens when an instance is
turned from master to slave. In that case it is more serious since this
will desynchronize data between slaves and masters. This other bug was
discovered as a side effect of thinking about the bug explained and
fixed in this commit, but will be fixed in a separated commit.
1. No need to set btype in processUnblockedClients(), since clients
flagged REDIS_UNBLOCKED should have it already cleared.
2. When putting clients in the unblocked clients list, clientsArePaused()
should flag them with REDIS_UNBLOCKED. Not strictly needed with the
current code but is more coherent.
When the list of unblocked clients were processed, btype was set to
blocking type none, but the client remained flagged with REDIS_BLOCKED.
When timeout is reached (or when the client disconnects), unblocking it
will trigger an assertion.
There is no need to process pending requests from blocked clients, so
now clientsArePaused() just avoid touching blocked clients.
Close#2467.
This commit moves the process of generating a new config epoch without
consensus out of the clusterCommand() implementation, in order to make
it reusable for other reasons (current target is to have a CLUSTER
FAILOVER option forcing the failover when no master majority is
reachable).
Moreover the commit moves other functions which are similarly related to
config epochs in a new logical section of the cluster.c file, just for
clarity.
Before we relied on the global cluster state to make sure all the hash
slots are linked to some node, when getNodeByQuery() is called. So
finding the hash slot unbound was checked with an assertion. However
this is fragile. The cluster state is often updated in the
clusterBeforeSleep() function, and not ASAP on state change, so it may
happen to process clients with a cluster state that is 'ok' but yet
certain hash slots set to NULL.
With this commit the condition is also checked in getNodeByQuery() and
reported with a identical error code of -CLUSTERDOWN but slightly
different error message so that we have more debugging clue in the
future.
Root cause of issue #2288.
Not perfect since The Solution IMHO is to have a DSL with a table of
configuration functions with type, limits, and aux functions to handle
the odd ones. However this hacky macro solution is already better and
forces to put limits in the range of numerical fields.
More field types to be refactored in the next commits hopefully.
Should be much faster, and regardless, the code is more obvious now
compared to generating a string just to get the return value of the
ll2stirng() function.
1. HVSTRLEN -> HSTRLEN. It's unlikely one needs the length of the key,
not clear how the API would work (by value does not make sense) and
there will be better names anyway.
2. Default is to return 0 when field is missing.
3. Default is to return 0 when key is missing.
4. The implementation was slower than needed, and produced unnecessary COW.
Related issue #2415.
1. Remove useless "cs" initialization.
2. Add a "select" var to capture a condition checked multiple times.
3. Avoid duplication of the same if (!copy) conditional.
4. Don't increment dirty if copy is given (no deletion is performed),
otherwise we propagate MIGRATE when not needed.
Less grays: more readable palette since usually we have a non linear
distribution of percentages and very near gray tones are hard to take
apart. Final part of the palette is gradient from yellow to red. The red
part is hardly reached because of usual distribution of latencies, but
shows up mainly when latencies are very high because of the logarithmic
scale, this is coherent to what people expect: red = bad.
The old version of SPOP with "count" argument used an API call of dict.c
which was actually designed for a different goal, and was not capable of
good distribution. We follow a different three-cases approach optimized
for different ratiion between sets and requested number of elements.
The implementation is simpler and allowed the removal of a large amount
of code.
Severan problems are addressed but still a few missing.
Since replication of this command was more complex than others since it
needs to replicate multiple SREM commands, an old API able to do this
was reused (it was taken inside the implementation since it was pretty
obvious soon or later that would be useful). The API was improved a bit
so that now a command may opt-out for the standard command replication
when the server.dirty counter is incremented, in order to "manually"
replicate what it wants.
--stat mode already used to reconnect automatically if the server is no
longer available. This is useful since this is an interactive mode used
for debugging, however the same applies to --latency and --latency-dist
modes, so now both use the reconnecting command execution as well.
The reconnection code was modified to use basic VT100 escape sequences
in order to play better with different kinds of output on the screen
when the reconnection happens, and to hide the reconnection attempt
output when finally the reconnection happens.
So far not able to find a color palette within the 256 colors which is
not confusing. However I believe it is a possible task, so will try
better later.
