[tarantool-patches] Re: [PATCH] replication: freeze join vclock before read view

[Vladislav Shpilevoy]

Hi! Thanks for the patch!

> diff --git a/src/box/relay.cc b/src/box/relay.cc
> index 74588cba7..ee23dd2aa 100644
> --- a/src/box/relay.cc
> +++ b/src/box/relay.cc
> @@ -298,6 +298,8 @@ relay_initial_join(int fd, uint64_t sync, struct vclock *vclock)
>  		relay_delete(relay);
>  	});
>  
> +	vclock_copy(vclock, &replicaset.vclock);
> +
>  	/* Freeze a read view in engines. */
>  	struct engine_join_ctx ctx;
>  	engine_prepare_join_xc(&ctx);
> @@ -312,8 +314,6 @@ relay_initial_join(int fd, uint64_t sync, struct vclock *vclock)
>  	if (wal_sync() != 0)
>  		diag_raise();
>  
> -	vclock_copy(vclock, &replicaset.vclock);
> -

I am not sure that you can just move the copy before wal sync.
As far as I understand, wal_sync() may commit some records and
change replicaset.vclock. As a result, after your patch
vclock will store not the latest vclock, won't it?

>  	/* Respond to the JOIN request with the current vclock. */
>  	struct xrow_header row;
>  	xrow_encode_vclock_xc(&row, vclock);