[Tarantool-patches] [PATCH vshard 1/2] router: drop wait_connected from master discovery

[Oleg Babin]

Thanks for your patch. Looks reasonable.

However I have one question. Before this patch if some

connection was down we perform wait_connected and in some

cases it could lead to successful reconnection. Currently we just skip

broken connections and don't try to reconnect.

Could it be a problem or we perform reconnection in another place?

What will be happen if all connections will be down?


On 04.12.2021 03:19, Vladislav Shpilevoy wrote:
> Master discovery tried to wait for connection establishment for
> the discovery timeout for each instance in the replicaset.
>
> The problem is that if one of replicas is dead, the discovery will
> waste its entire timeout on just this waiting. For all the
> requests sent to connected replicas after this one it will have 0
> timeout and won't properly wait for their results.
>
> For example, this is how master discovery could work:
>
>      send requests:
>          replica1 wait_connected + send,
>          replica2 wait_connected fails on timeout
>          replica3 wait_connected works if was connected + send
>
>      collect responses:
>          replica1 wait_result(0 timeout)
>          replica2 skip
>          replica3 wait_result(0 timeout)
>
> The entire timeout was wasted on 'replica2 wait_connected' during
> request sending. Replica1 result could be delivered fine because
> it was in progress while replica2 was waiting. So having 0 timeout
> in it is not a problem. It had time to be executed. But replica3's
> request has very few chances to be delivered in time. It was just
> sent and is collected almost immediately.
>
> The worst case is when the first replica is dead. Then it is very
> likely neither of requests will be delivered. Due to all result
> wait timeouts being 0.
>
> Although there is a certain chance that the next requests will be
> extra quick, so writing a stable test for that does not seem
> possible.
>
> The bug was discovered while working on #288. For its testing it
> was needed to stop one instance and master_discovery test started
> failing.
> ---
>   vshard/replicaset.lua | 6 +-----
>   1 file changed, 1 insertion(+), 5 deletions(-)
>
> diff --git a/vshard/replicaset.lua b/vshard/replicaset.lua
> index 55028bd..174c761 100644
> --- a/vshard/replicaset.lua
> +++ b/vshard/replicaset.lua
> @@ -682,11 +682,7 @@ local function replicaset_locate_master(replicaset)
>       local replicaset_uuid = replicaset.uuid
>       for replica_uuid, replica in pairs(replicaset.replicas) do
>           local conn = replica.conn
> -        timeout, err = netbox_wait_connected(conn, timeout)
> -        if not timeout then
> -            last_err = err
> -            timeout = deadline - fiber_clock()
> -        else
> +        if conn:is_connected() then
>               ok, f = pcall(conn.call, conn, func, args, async_opts)
>               if not ok then
>                   last_err = lerror.make(f)