Deadlock with redis

[thorstenfleischmann]

I experience deadlocks when having a lot of parallel requests.

With set $session_redis_pool_size 1 I can reproduce it always with just two requests.

The code is basically:

session:open()
dosomething()
session:close()

With some debug logs I think I figured out the reason.

1. request 1 opens/locks session
2. request 1 "open" puts connect back to pool ("set_keepalive")
3. requests 1 does something
4. requests 2 comes in and opens connection to redis pool. session is locked so it trys to aquire lock which takes time to time out
5. request 1 did something and wants to close session. session:close cannot get a connection to unlock because requests 2 blocks the session waiting for the lock

Is there a good way to prevent this?

I am using lua resty session 3.

With a bigger pool size there are factors which makes the deadlock more or less likely. It looks like http2 is a factor since there can be more parallel client connections.

[thorstenfleischmann]

I found a possible solution, but it requires code change in storage/redis.lua at function storage.lock(key)

        -- put connection back to pool
        self:set_keepalive()

        sleep(waittime)

        -- reconnect
        local ok, err = self:connect()
        if not ok then
            return nil, err
        end

This puts the connection during spin lock wait back to the connection pool. This might introduce some overhead during waiting for a lock but it eliminates the deadlocks.

[bungle]

@thorstenfleischmann,

The 4.x uses only lockless mode. With 3.x you can use regenerate strategy:
https://github.com/bungle/lua-resty-session/tree/v3.10?tab=readme-ov-file#pluggable-session-strategies

[thorstenfleischmann]

Regenerate strategy does not really solve this. My example uses session read only which does not generate a new session.

Unfortunately locking was crucial for my use case.