refactor client tracking, fix atomicity, squashing and multi/exec

[kostasrim]

• add partial support for CLIENT CACHING TRUE (only to be used with TRACKING OPTIN)
• add OPTIN to CLIENT TRACKING command
• refactor client tracking to respect transactional atomicity
• fixed multi/exec and disabled squashing with client tracking
• add tests

Resolves #2969, #2971, #2997, #2998

P.s. All tests in rueidis TestSingleClientIntegration pass except pub/sub because we don't yet support it see #3001

[kostasrim]

@dranikpg this with the changes in InvokeCmd seemed to be the most non intrusive way (to comply with the requirements of the state machine)

[romange]

1. I would like to understand why tracking requires transaction semantics (an example will be fine)
2. why cid_ is not enough and we need invoke_cid_ ?

[dranikpg]

It would suggest to do it like we handle blocking commands - once it finished and manually from RunSquashedCb

dragonfly/src/server/transaction.cc

Lines 638 to 640 in a95419b

	if (auto* bcontroller = shard->blocking_controller(); bcontroller) {
	if (awaked_prerun || was_suspended) {
	bcontroller->FinalizeWatched(GetShardArgs(idx), this);

So it becomes if (concluding || (multi && multi_->concluding)) Track(this)

Now you don't need invoke_cid, etc there as well

[kostasrim]

I would like to understand why tracking requires transaction semantics (an example will be fine)

Because invalidation messages must be sent before the transaction concludes. Otherwise, we might accidentally skip them. An example would be:

>> CLIENT TRACKING ON
>> GET FOO
>> SET FOO BAR 
>> GET FOO
>> SET FOO BAR
>> GET FOO ---------> might miss Invalidation message

A valid execution would be once we call the first SET we will send an invalidation message as a separate transaction. Now before that even starts/concludes, the GET that follows will get executed first and it will itself issue a separate transaction to send an invalidation message. Now the problem here is, that once we send an invalidation message we remove the key from the tracking map (since we only send invalidation messages once until the key is reread). Then the second invalidation transaction won't work because the key no longer exists in the map and we will never get that second invalidation message.

[kostasrim]

That;s for squashing :)

[dranikpg]

If you access conn_state, don't make it a function on conn_cntx

then you can just use conn_state, you can make it mutable or add a new member like conn

dragonfly/src/server/main_service.cc

Lines 214 to 215 in a95419b

	if (cntx->conn_state.squashing_info)
	cntx = cntx->conn_state.squashing_info->owner;

[romange]

why do you need to call it here?

[romange]

seems that what you really want is to know if you are in the middle of EXEC execution and not multi.

[dranikpg]

We store so much fragile info that needs to be updated everywhere... seqnums would solve all this

[romange]

nit: UdatePrevAndLastCommand name describes the implementation of this function. What it does is advancing the state. So I think it's better call it Tick or Advance or Update

[romange]

rename: executing_command_ to track_next_cmd_

[kostasrim]

but the track_next_cmd_ seems misleading since it implies it's the next command. executing_command_ is the command we currently execute in InvokeCmd flow and prev_command_ is the command before it. So:

>> GET FOO ----> prev_command
>> GET BAR ----> current_command

[romange]

rename prev_command_ to track_current_cmd_

[dranikpg]

It looks somewhat over engineered to me 😅 We have similar semantics in blocking commands - only that we subscribe with specific commands and not with any.

1. Let's use squashing_info instead of adding a parent field to ConnectionContext or let's use that field for everything - there should be one way of doing things with proper comments, so nobody adds yet a third
2. I'd still suggest to add numbers to commands, because UpdatePrevAndLastCommand() appears in many places and we update three whole fileds: prev, executing, multi. The track command can just store its number and we don't have to update much more
3. Track() should be called when we conclude or finish the current multi command, currently we call it for every hop. Not that there are multi-hop read commands, but I think it belongs to all other management code. Invoke-cid should also not be needed with that

[dranikpg]

It would suggest to do it like we handle blocking commands - once it finished and manually from RunSquashedCb

dragonfly/src/server/transaction.cc

Lines 638 to 640 in a95419b

	if (auto* bcontroller = shard->blocking_controller(); bcontroller) {
	if (awaked_prerun || was_suspended) {
	bcontroller->FinalizeWatched(GetShardArgs(idx), this);

So it becomes if (concluding || (multi && multi_->concluding)) Track(this)

Now you don't need invoke_cid, etc there as well

[dranikpg]

If you access conn_state, don't make it a function on conn_cntx

then you can just use conn_state, you can make it mutable or add a new member like conn

dragonfly/src/server/main_service.cc

Lines 214 to 215 in a95419b

	if (cntx->conn_state.squashing_info)
	cntx = cntx->conn_state.squashing_info->owner;

[dranikpg]

We store so much fragile info that needs to be updated everywhere... seqnums would solve all this

[dranikpg]

See previous comment on whether we can keep this in conn_state

[romange]

I think this comment can be improved. Explain how it happens that DCHECK(dispatch_q_.empty()); fails after the migration, "rueidis SingleIntegrationTest" is irrelevant here

[romange]

this function now is called within the Shard. We have a convention to name it as xxxOnShard to stress it.

[romange]

I believe the code can be moved to the coordinator thread, i.e. when we prepare the transaction. the result can be kept as a boolean within transaction. Nothing here is relevant to the shard or the transaction state. we have Transaction::coordinator_state_ mask that can be used for this.

[romange]

actually, see my other comment.

[kostasrim]

Well, info.ShouldTrackeys() should be called after a command is invoked because that very command might be the client tracking command.

The callback approach is neat for removing the ConnectionContext alltogether from transaction and in fact this is the only change needed (no fancy boolean or coordinator mask) because we just pass a callback which just captures the context :)

[romange]

we do not need this move now

[romange]

looks good, left final minor comments.

[romange]

why trans->SetTrackingCallback({}); ? can you add a comment please?

[romange]

a better interface for tracking callback would be to accept Transaction* and EngineShard* as its arguments, so the transaction can just call it with these variables instead of capturing them here.

[romange]

can you please add a short comment what/why we do here?

[romange]

do we still need to remove const here?

[romange]

!!!