* Default reconnect strategy uses exponential backoff and jitter
Both are recommended parts of client reconnect strategies to prevent
thundering herd problems when many clients lose their connection at once
(for example, during a Redis upgrade).
* Move default retry strategy to constant
* Plain english explanation of default 'socket.reconnectStrategy'
* Extract default connect strategy into helper function
* redis client socket changes needed for sentinel
* Sentinel Implementation [EXPERIMENTAL]
* add pooling
* improve typing with SENTINEL_ client members
* cleanup - remove unused comments / commented code
* small sendCommand change + revert change to tsconfig
* add more sentinel commands needed for testing.
* lots of fixups and a reasonable first pass test suite
* add a timer option to update topology in background
+ don't need both sentinel client and pubsubclient
+ nits
* format all the things
* more progress
* small cleanup
* try to group promises together to minimize the internal await points
* redo events, to keep a single topology event to listen on
* nits + readme
* add RedisSentinelFactory to provide lower level access to sentinel
* nit
* update
* add RedisSentinelClient/Type for leased clients
returned by aquire()
used by function passed to use()
* add self for private access + improve emitting
* nit
* nits
* improve testing
- improve steady state waiting between tests
- get masternode from client, not from sentinels themselves (not consistent and then client isn't changing as we expect
- provide extensive logging/tracing on test errors
- provide a very low impact tracing mechanism withinthe code that only really impacts code when tracing is in use.
* ismall nit for typing
* bunch of changes
- harden testing
- don't use sentinel[0] for debug error dump as could be downed by a test
- increase time for sentinel down test to 30s (caused a long taking failover)
- add client-error even / don't pass throuh client errors as errors option for pubsub proxy
- when passing through cient errors as error events, dont pass the event, but the Error object, as only Error objects are supposed to be on 'error'
-
* improve pub sub proxy.
save the refference to all channel/pattern listeners up front on creation, dont hve to fetch the object each time, as it doesn't change.
removes race condition between setting up the listener and the pub sub node going down and being recreated.
* wrap the passed through RedisClient error to make clear where its coming from.
* refactor sentinel object / factory tests apart
* harden tests a little bit more
* add pipeline test
* add scripts/function tests + fixups / cleanups to get them to work
* change to use redis-stack-server for redis nodes to enable module testing
* fix test, forgot to return in use function with module
* rename test
* improve tests to test with redis/sentinel nodes with and withput passwords
this tests that we are handling the nodeClientOptions and sentinelClientOptions correctly
* cleanup for RedisSentinel type generic typing in tests
* remove debugLog, just rely on traace mechanism
* added multi tests for script/function/modules
* don't emit errors on lease object, only on main object
* improve testing
* extract out common code to reduce duplication
* nit
* nits
* nit
* remove SENTINEL_... commands from main client, load them via module interface
* missed adding RedisSentinelModule to correct places in RedisSentinelFactory
* nits
* fix test logging on error
1) it takes a lot of time now, so needs larger timeout
2) docker logs can be large, so need to increase maxBuffer size so doesn't error (and break test clean up)
* invalidate watches when client reconnects
+ provide API for other wrapper clients to also create invalid watch states programatically.
Reasoning: if a user does a WATCH and then the client reconnects, the watch is no longer active, but if a user does a MULTI/EXEC after that, they wont know, and since the WATCH is no longer active, the request has no protection.
The API is needed for when a wrapper client (say sentinel, cluster) might close the underlying client and reopen a new one transparently to the user. Just like in the reconnection case, this should result in an error, but its up to the wrapping client to provide the appropriate error
* remove WATCH and UNWATCH command files, fix WATCH and UNWATCH return type, some more cleanups
* missing file in last commit :P
* support for custom message in `WatchError`
* setDirtyWatch
* update watch docs
* fixes needed
* wip
* get functions/modules to work again
self -> _self change
* reuse leased client on pipelined commands.
though I realize this implementation, really only works after the first write command.
unsure this is worth it.
* test tweaks
* nit
* change how "sentinel" object client works, allow it to be reserved
no more semaphore type counting
* review
* fixes to get more tests to pass
* handle dirtyWatch and watchEpoch in reset and resetIfDirty
* "fix", but not correct, needs more work
* fix pubsub proxy
* remove timeout from steadyState function in test, caused problems
* improve restarting nodes
* fix pubsub proxy and test
---------
Co-authored-by: Leibale Eidelman <me@leibale.com>
