Re: [Tarantool-patches] [RFC v3 2/3] applier: send first row's WAL time in the applier_writer_f

[Cyrill Gorcunov via Tarantool-patches <tarantool-patches@dev.tarantool.org>]

On Fri, Apr 30, 2021 at 10:49:10PM +0200, Vladislav Shpilevoy wrote:
> > 
> > Same time the apply_plain_tx uses asynchronous WAL write completion
> > and at moment when the write procedure is finished the applier might
> > be removed from replicaset already thus we use applier's instance
> 
> Did you mean instance id?

Yes.

> > to lookup if it is still alive.
> > 
> > The calculation of the downstream lag itself lag will be addressed
> 
> One of the 'lag' words is redundant.

Thanks for catching!

> >  
> > +/** Applier WAL related statistics */
> > +struct awstat {
> 
> 1. Please, lets avoid such hard contractions. At first I
> thought you decided to do something related to AWS when saw
> this name.
> 
> 	struct applier_lag
> 
> would be just fine.

Sure, will update.

> 
> > +	uint32_t instance_id;
> > +	double first_row_tm;
> > +};
> > +
> > +static void
> > +awstat_update(struct awstat *awstat)
> > +{
> > +	/* Ignore if not needed */
> > +	if (awstat->instance_id == 0)
> > +		return;
> 
> 2. Why did you even allocate this stat if it is not needed?
> Maybe it would be better to have it NULL then and check
> for NULL? AFAIU these are the initial and final join cases.

They are allocated on the stack together with synchro_entry,
so there is no penalty. After your comment I've changed it to

struct synchro_entry {
	/** Request to process when WAL write is done. */
	struct synchro_request *req;
	/** Fiber created the entry. To wakeup when WAL write is done. */
	struct fiber *owner;
	/** WAL bound statistics. */
	struct applier_lag *applier_lag;
	/**
	 * The base journal entry. It has unsized array and then must be the
	 * last entry in the structure. But can workaround it via a union
	 * adding the needed tail as char[].
	 */
	union {
		struct journal_entry base;
		char base_buf[sizeof(base) + sizeof(base.rows[0])];
	};
};

and then allocate applier_lag on the stack and assign a pointer if needed.

static int
apply_synchro_row(struct applier *applier, struct xrow_header *row,
		  bool use_awstat)
{
	...
	struct applier_lag applier_lag;
	struct synchro_entry entry;
	...
	if (use_awstat) {
		applier_lag.instance_id = applier->instance_id;
		applier_lag.first_row_tm = row->tm;

		entry.applier_lag = &applier_lag;
	} else {
		entry.applier_lag = NULL;
	}
}

Strictly speaking there is no difference.

> 
> Did you try the way I proposed about waiting for all applier's
> WAL writes to end in applier_stop()? Does it look worse? After
> the fiber stop and wal_sync() it would be safe to assume there
> are no WAL writes in fly from this applier. But I don't know if
> it would look better.

I thought about it alot. And you know, I don't really like what we
are to implement

 - currently applier_stop() doesn't wait the journal to finish its
   write. The main applier reader is spinning in !fiber_is_cancelled()
   cycle in a polling way while applier tries to read new data from the
   remote relay peer. If peer doesn't reply for some reason then we throw
   an exception which is catched by a caller code, and the caller tries
   to iterate new cycle testing if fiber is cancelled.

   In case if reconfiguration happens (and timeouts are set to default)
   then we try to prune old appliers calling fiber_join on them which
   means this fiber_join won't exit until new fiber_is_cancelled iteration
   get processed. With default configs it means this could take up to
   replication_disconnect_timeout(), ie 1 second by default.

   In turn if we bound to journal write completion then this gonna be
   uncontrollable because journal may hang as long as it wish and we
   continue spinning in a waiting cycle

 - in applier_stop() we will have to implement some kind of reference
   counting, which would be modified on journal completion and i think
   this makes code even more complex, since we have to add some additional
   logic when applier is allowed to cancel.

> > +
> > +	/*
> > +	 * Write to WAL happens in two contexts: as
> > +	 * synchronous writes and as asynchronous. In
> > +	 * second case the applier might be already
> > +	 * stopped and removed.> +	 */
> > +	struct replica *r = replica_by_id(awstat->instance_id);
> > +	if (r == NULL && r->applier == NULL)
> 
> 3. There is another idea - store the timestamp in struct replica.
> Then it is -1 dereference. Although you would need to call
> replica_by_id() before each ACK, but one ACK covers multiple
> transactions and it would mean less lookups than now.

I like this idea, letme try to implement it.

> > +		return;
> > +
> > +	r->applier->first_row_wal_time = awstat->first_row_tm;
> 
> 4. In case there was a batch of transactions written to WAL,
> the latest one will override the timestamp of the previous ones and
> this would make the lag incorrect, because you missed the older
> transactions. Exactly like when you tried to take a timestamp of
> the last row instead of the first row, but in a bigger scope.
> Unless I missed something.

I'm not sure I follow you here. Say we have a batch of transactions.
The update happens on every journal_entry completion, if several
entries are flushed then completion is called in ordered way (one
followed by another). The update happens in same tx thread where
appliers are running which means acks sending procedure is ordered
relatively to updates call. Thus we may have situation where we
complete first entry then we either send it in ack message either
update to a new value and only then send ack. As far as I understand
there might be a gap in fibers scheduling before several journal entries
get completed. Thus the lag calculated on relay side will be bigger for
some moment but on next ack will shrink to latest entry for writtent
journal entry. Because relay uses current time minus time value
obtained from applier's ack. Hopefully I didn't miss something either.

> Probably you need to assign a new timestamp only when the old
> one is not 0, and reset it to 0 on each sent ACK. Don't know.

Gimme some time to investigate this vague moment.

> 
> > @@ -817,6 +852,12 @@ apply_synchro_row(struct xrow_header *row)
> >  			     apply_synchro_row_cb, &entry);
> >  	entry.req = &req;
> >  	entry.owner = fiber();
> > +	if (use_awstat) {
> 
> 5. You don't really need this flag. Because during joins applier's
> instance ID should be zero anyway. Therefore you would assign
> stat.instance_id = 0 in this case regardless of the flag.
> 
> Also this means you don't need the entire applier struct. You only
> need the instance_id as an argument. I am saying this because these
> functions apply_* didn't not depend on a concrete applier object
> probably exactly because it could be deleted. Not having an applier
> pointer in their arguments could be intentional.

Sounds reasonable, thanks!

> > @@ -93,6 +93,11 @@ struct applier {
> >  	ev_tstamp last_row_time;
> >  	/** Number of seconds this replica is behind the remote master */
> >  	ev_tstamp lag;
> > +	/**
> > +	 * WAL time of first applied row in a transaction.
> > +	 * For relay statistics sake.
> 
> 6. It is not a first applied row on the whole. Only for the latest
> transaction written to WAL.

+1

Thanks a huge for comments, Vlad!