This is apparent when invalidation messages are sent on the same connection, as they may be received before or after the mutating command's reply.

For example, HSET delivers the reply first:

hget h f
_
hset h f 1
:1
>2
$10
invalidate
*1
$1
h

On the other hand, HDEL will deliver the invalidation first:

hget h f
$1
1
hdel h f
>2
$10
invalidate
*1
$1
h
:1

It seems the order of producing the reply, calling signalModifiedKey and creating a keyspace notification (which is also affected) is arbitrary. I'm not sure if the order makes a difference here but it should probably be consistent. @redis/core-team any thoughts?

Thanks @michael-grunder for bringing this up (and fixing hiredis to handle that)!