[Tarantool-patches] [PATCH 5/6] replication: use 'score' to find a join-master

[Serge Petrenko]

05.06.2021 02:37, Vladislav Shpilevoy пишет:
> The patch refactors the algorithm of finding a join-master (in
> replicaset_find_join_master()) to use scores instead of multiple
> iterations with different criteria.
>
> The original code was relatively fine as long as it had only
> one parameter to change - whether should it skip
> `box.cfg{read_only = true}` nodes.
>
> Although it was clear that it was "on the edge" of acceptable
> complexity due to a second non-configurable parameter whether a
> replica is in read-only state regardless of its config.
>
> It is going to get more complicated when the algorithm will take
> into account the third parameter whether an instance is
> bootstrapped.
>
> Then it should make decisions like "among bootstrapped nodes try
> to prefer instances not having read_only=true, and not being in
> read-only state". The easiest way to do so is to use
> scores/weights incremented according to the instance's parameters
> matching certain "good points".
>
> Part of #5613

LGTM.

> ---
>   src/box/replication.cc | 62 ++++++++++++++++--------------------------
>   1 file changed, 23 insertions(+), 39 deletions(-)
>
> diff --git a/src/box/replication.cc b/src/box/replication.cc
> index 990f6239c..d33e70f28 100644
> --- a/src/box/replication.cc
> +++ b/src/box/replication.cc
> @@ -960,71 +960,55 @@ replicaset_next(struct replica *replica)
>    * replicas, choose a read-only replica with biggest vclock
>    * as a leader, in hope it will become read-write soon.
>    */
> -static struct replica *
> -replicaset_round(bool skip_ro)
> +struct replica *
> +replicaset_find_join_master(void)
>   {
>   	struct replica *leader = NULL;
> +	int leader_score = -1;
>   	replicaset_foreach(replica) {
>   		struct applier *applier = replica->applier;
>   		if (applier == NULL)
>   			continue;
>   		const struct ballot *ballot = &applier->ballot;
> -		/**
> -		 * While bootstrapping a new cluster, read-only
> -		 * replicas shouldn't be considered as a leader.
> -		 * The only exception if there is no read-write
> -		 * replicas since there is still a possibility
> -		 * that all replicas exist in cluster table.
> -		 */
> -		if (skip_ro && ballot->is_ro_cfg)
> -			continue;
> -		if (leader == NULL) {
> -			leader = replica;
> -			continue;
> -		}
> -		const struct ballot *leader_ballot = &leader->applier->ballot;
> +		int score = 0;
>   		/*
> -		 * Try to find a replica which has already left
> -		 * orphan mode.
> +		 * Prefer instances not configured as read-only via box.cfg, and
> +		 * not being in read-only state due to any other reason. The
> +		 * config is stronger because if it is configured as read-only,
> +		 * it is in read-only state for sure, until the config is
> +		 * changed.
>   		 */
> -		if (ballot->is_ro && !leader_ballot->is_ro)
> +		if (!ballot->is_ro_cfg)
> +			score += 5;
> +		if (!ballot->is_ro)
> +			score += 1;
> +		if (leader_score < score)
> +			goto elect;
> +		if (score < leader_score)
>   			continue;
> +		const struct ballot *leader_ballot;
> +		leader_ballot = &leader->applier->ballot;
>   		/*
>   		 * Choose the replica with the most advanced
>   		 * vclock. If there are two or more replicas
>   		 * with the same vclock, prefer the one with
>   		 * the lowest uuid.
>   		 */
> -		int cmp = vclock_compare_ignore0(&ballot->vclock,
> -						 &leader_ballot->vclock);
> +		int cmp;
> +		cmp = vclock_compare_ignore0(&ballot->vclock,
> +					     &leader_ballot->vclock);
>   		if (cmp < 0)
>   			continue;
>   		if (cmp == 0 && tt_uuid_compare(&replica->uuid,
>   						&leader->uuid) > 0)
>   			continue;
> +	elect:
>   		leader = replica;
> +		leader_score = score;
>   	}
>   	return leader;
>   }
>   
> -struct replica *
> -replicaset_find_join_master(void)
> -{
> -	bool skip_ro = true;
> -	/**
> -	 * Two loops, first prefers read-write replicas among others.
> -	 * Second for backward compatibility, if there is no such
> -	 * replicas at all.
> -	 */
> -	struct replica *leader = replicaset_round(skip_ro);
> -	if (leader == NULL) {
> -		skip_ro = false;
> -		leader = replicaset_round(skip_ro);
> -	}
> -
> -	return leader;
> -}
> -
>   struct replica *
>   replica_by_uuid(const struct tt_uuid *uuid)
>   {

-- 
Serge Petrenko