fix(plugins) don't call redis:select when not necessary #3973
Conversation
In plugin rate-limiting and response-ratelimiting, a redis:select is called when connection is reused, even though the current configured redis database is 0, to clear the previously selected database. This will be problem when for some use case like twemproxy where select is not supported, and sending `select` call will cause the connection to be closed. The patch used a host+port+database based redis connection pool, so that we always know the current connection is using correct database.
| -- return a special pool name only if redis_database is set to non-zero | ||
| -- otherwise use the default pool name host:port | ||
| if conf.redis_database ~= 0 then | ||
| local key = fmt("%s:%d:%d", conf.redis_port, conf.redis_port, conf.redis_database) |
There was a problem hiding this comment.
The second argument here should be conf.redis_host
There was a problem hiding this comment.
Additionally, prior art has proven that a simple concat operation could perform as well or better than string.format, we should maybe follow this new rule of thumb.
| @@ -136,7 +145,8 @@ return { | |||
| increment = function(conf, limits, identifier, current_timestamp, value) | |||
| local red = redis:new() | |||
| red:set_timeout(conf.redis_timeout) | |||
| local ok, err = red:connect(conf.redis_host, conf.redis_port) | |||
| local ok, err = red:connect(conf.redis_host, conf.redis_port, | |||
| { pool = get_redis_pool_name(conf) }) | |||
There was a problem hiding this comment.
We could reuse this table (if declared at the top chunk level) and just override its pool property, to save some GC overhead, and some function call overhead as well:
sock_opts.pool = get_redis_pool_name(conf) -- could even be inlined
thibaultcha
added
the
pr/changes requested
|
Nice catch :) Meant to request some changes, but mis-clicked. Additionally, note that the linter is failing in CI. |
|
@thibaultcha Thank you for reviewing this, I've addressed the comments and CI is green now. Could you take another look when you got chance? |
| if not ok then | ||
| ngx_log(ngx.ERR, "failed to change Redis database: ", err) | ||
| return nil, err | ||
| local ok, err = red:select(conf.redis_database or 0) |
There was a problem hiding this comment.
probably dont need the or 0 here?
|
LGTM, however we should wait until our Travis CI issues are resolved before merging this. |
|
Hi @thibaultcha, do we have an update on the CI issue? |
|
@fffonion It should be fine if you rebase this branch on the latest master. |
|
Running green on |
hishamhm
deleted the
fix/plugins-dont-call-redis-select-when-not-necessary
branch
Summary
In plugin rate-limiting and response-ratelimiting, a redis:select is
called when connection is reused, even though the current configured
redis database is 0, to clear the previously selected database. This
will be problem for some use case like twemproxy where
selectisnot supported, and sending
selectcall will cause the connection tobe closed.
The patch used a host+port+database based redis connection pool, so that
we always know the current connection is using correct database. The
behaviour of redis:auth only once is not changed.
Full changelog
selectcallIssues resolved
Related to #3293