[tarantool-patches] Re: [PATCH 1/3] Update only vshard part of a cfg on reload
Vladislav Shpilevoy
v.shpilevoy at tarantool.org
Wed Aug 1 21:43:35 MSK 2018
Thanks for the patch! See 4 comments below.
On 31/07/2018 19:25, AKhatskevich wrote:
> Box cfg could have been changed by a user and then overridden by
> an old vshard config on reload.
>
> Since that commit, box part of a config is applied only when
> it is explicitly passed to a `cfg` method.
>
> This change is important for the multiple routers feature.
> ---
> vshard/cfg.lua | 54 +++++++++++++++++++++++++------------------------
> vshard/router/init.lua | 18 ++++++++---------
> vshard/storage/init.lua | 53 ++++++++++++++++++++++++++++--------------------
> 3 files changed, 67 insertions(+), 58 deletions(-)
>
> diff --git a/vshard/cfg.lua b/vshard/cfg.lua
> index bba12cc..8282086 100644
> --- a/vshard/cfg.lua
> +++ b/vshard/cfg.lua
> @@ -230,48 +230,50 @@ local non_dynamic_options = {
> 'bucket_count', 'shard_index'
> }
>
> +--
> +-- Deepcopy a config and split it into vshard_cfg and box_cfg.
> +--
> +local function split_cfg(cfg)
> + local vshard_field_map = {}
> + for _, field in ipairs(cfg_template) do
> + vshard_field_map[field[1]] = true
> + end
1. vshard_field_map does not change ever. Why do you build it
on each cfg? Please, store it in a module local variable like
cfg_template. Or refactor cfg_template and other templates so
they would be maps with parameter name as a key - looks like
the most suitable solution.
> + local vshard_cfg = {}
> + local box_cfg = {}
> + for k, v in pairs(cfg) do
> + if vshard_field_map[k] then
> + vshard_cfg[k] = table.deepcopy(v)
> + else
> + box_cfg[k] = table.deepcopy(v)
> + end
> + end
> + return vshard_cfg, box_cfg
> +end
> +
> diff --git a/vshard/storage/init.lua b/vshard/storage/init.lua
> index 102b942..75f5df9 100644
> --- a/vshard/storage/init.lua
> +++ b/vshard/storage/init.lua
> @@ -1500,13 +1500,17 @@ end
> --------------------------------------------------------------------------------
> -- Configuration
> --------------------------------------------------------------------------------
> +-- Private (not accessible by a user) reload indicator.
> +local is_reload = false
2. Please, make this variable be parameter of storage_cfg and wrap public
storage.cfg with a one-liner:
storage.cfg = function(cfg, uuid) return storage_cfg(cfg, uuid, false) end
I believe/hope you understand that such way to pass parameters, via global
variables, is flawed by design.
> @@ -1553,18 +1557,19 @@ local function storage_cfg(cfg, this_replica_uuid)
> --
> -- If a master role of the replica is not changed, then
> -- 'read_only' can be set right here.
> - cfg.listen = cfg.listen or this_replica.uri
> - if cfg.replication == nil and this_replicaset.master and not is_master then
> - cfg.replication = {this_replicaset.master.uri}
> + box_cfg.listen = box_cfg.listen or this_replica.uri
> + if box_cfg.replication == nil and this_replicaset.master
> + and not is_master then
3. Broken indentation.
> + box_cfg.replication = {this_replicaset.master.uri}
> else
> - cfg.replication = {}
> + box_cfg.replication = {}
> end
> if was_master == is_master then
> - cfg.read_only = not is_master
> + box_cfg.read_only = not is_master
> end
> if type(box.cfg) == 'function' then
> - cfg.instance_uuid = this_replica.uuid
> - cfg.replicaset_uuid = this_replicaset.uuid
> + box_cfg.instance_uuid = this_replica.uuid
> + box_cfg.replicaset_uuid = this_replicaset.uuid
> else
> local info = box.info
> if this_replica_uuid ~= info.uuid then
> @@ -1607,9 +1614,10 @@ local function storage_cfg(cfg, this_replica_uuid)
> local_on_master_enable_prepare()
> end
>
> - local box_cfg = table.copy(cfg)
> - lcfg.remove_non_box_options(box_cfg)
> - local ok, err = pcall(box.cfg, box_cfg)
> + local ok, err = true, nil
> + if not xis_reload then
> + ok, err = pcall(box.cfg, box_cfg)
> + end
4. The code below (if not ok then ...) can be moved inside
'if not is_reload' together with 'local ok, err' declaration.
Please, do.
> while M.errinj.ERRINJ_CFG_DELAY do
> lfiber.sleep(0.01)
> end
More information about the Tarantool-patches
mailing list