[tarantool-patches] Re: [PATCH 1/4] Fix races related to object outdating

[Alex Khatskevich]

On 01.08.2018 15:36, Vladislav Shpilevoy wrote:
> Hi! Thanks for the patch! See 2 comments below.
>
>> diff --git a/vshard/router/init.lua b/vshard/router/init.lua
>> index 142ddb6..1a0ed2f 100644
>> --- a/vshard/router/init.lua
>> +++ b/vshard/router/init.lua
>> @@ -88,15 +94,18 @@ local function bucket_discovery(bucket_id)
>>      log.verbose("Discovering bucket %d", bucket_id)
>>      local last_err = nil
>>      local unreachable_uuid = nil
>> -    for uuid, replicaset in pairs(M.replicasets) do
>> -        local _, err =
>> -            replicaset:callrw('vshard.storage.bucket_stat', 
>> {bucket_id})
>> -        if err == nil then
>> -            bucket_set(bucket_id, replicaset)
>> -            return replicaset
>> -        elseif err.code ~= lerror.code.WRONG_BUCKET then
>> -            last_err = err
>> -            unreachable_uuid = uuid
>> +    for uuid, _ in pairs(M.replicasets) do
>> +        -- Handle reload/reconfigure.
>> +        replicaset = M.replicasets[uuid]
>
> 1. Please, explain, how is it possible, that before this line
> M.replicasets[uuid] can become nil. You iterate here over
> M.replicasets and on the previous line in '_' you have
> stored replicaset. How can here 'replicaset' differ from '_'?
>
> It has nothing in common with 'reload/reconfigure' case since
> you always iterate over M.replicasets - the most actual
> list of replicasets. Maybe you thought that pairs() stores
> its first argument into a temporary variable but looks like
> it is not. I checked it with a simple test:
>
>     a = {}
>     a.objs = {1, 2, 3, 4, 5}
>     for k, v in pairs(a.objs) do
>         print(a.objs[k])
>         if k == 2 then
>             a.objs = {6, 7, 8, 9, 10}
>         end
>     end
>
>     Output:
>     1
>     2
>     8
>     9
>     10
>     ---
>     ...
Thanks. For some reason I was sure that pairs caches the table.
Fixed.
>> +        if replicaset then
>> +            local _, err =
>> +                replicaset:callrw('vshard.storage.bucket_stat', 
>> {bucket_id})
>> +            if err == nil then
>> +                return bucket_set(bucket_id, replicaset.uuid)
>> +            elseif err.code ~= lerror.code.WRONG_BUCKET then
>> +                last_err = err
>> +                unreachable_uuid = uuid
>> +            end
>>          end
>>      end
>>      local err = nil
>> @@ -513,27 +522,28 @@ local function router_cfg(cfg)
>>      end
>>      box.cfg(box_cfg)
>>      log.info("Box has been configured")
>> -    M.connection_outdate_delay = cfg.connection_outdate_delay
>> -    M.total_bucket_count = total_bucket_count
>> -    M.collect_lua_garbage = collect_lua_garbage
>> -    M.current_cfg = new_cfg
>>      -- Move connections from an old configuration to a new one.
>>      -- It must be done with no yields to prevent usage both of not
>>      -- fully moved old replicasets, and not fully built new ones.
>> -    lreplicaset.rebind_replicasets(new_replicasets, M.replicasets,
>> -                                   M.connection_outdate_delay)
>> -    M.replicasets = new_replicasets
>> +    lreplicaset.rebind_replicasets(new_replicasets, M.replicasets)
>>      -- Now the new replicasets are fully built. Can establish
>>      -- connections and yield.
>>      for _, replicaset in pairs(new_replicasets) do
>>          replicaset:connect_all()
>>      end
>> +    lreplicaset.wait_masters_connect(new_replicasets)
>> +    lreplicaset.outdate_replicasets(M.replicasets, 
>> cfg.connection_outdate_delay)
>> +    M.connection_outdate_delay = cfg.connection_outdate_delay
>> +    M.total_bucket_count = total_bucket_count
>> +    M.collect_lua_garbage = collect_lua_garbage
>> +    M.current_cfg = cfg
>> +    M.replicasets = new_replicasets
>>      -- Update existing route map in-place.
>> -    for bucket, rs in pairs(M.route_map) do
>> +    local old_route_map = M.route_map
>> +    M.route_map = {}
>> +    for bucket, rs in pairs(old_route_map) do
>>          M.route_map[bucket] = M.replicasets[rs.uuid]
>
> 2. Why do you need to save old_route_map into a
> separate variable? You can update M.route_map in place
> like it was done before. It is not? You fill the new
> route_map with exactly the same keys (maybe with different
> values, but it does not affect 'for' iteration). Moreover,
> when you create a new route_map instead of resetting the
> old, you double the memory. For huge bucket count it can
> be noticeable.
Yes there is no point in that change.
Fixed.
>>      end
>> -
>> -    lreplicaset.wait_masters_connect(new_replicasets)
>>      if M.failover_fiber == nil then
>>          lfiber.create(util.reloadable_fiber_f, M, 'failover_f', 
>> 'Failover')
>>      end