[PATCH 2/2] schema: expose space_mt and index_mt on `box.schema` table

[Vladimir Davydov]

On Sun, Apr 01, 2018 at 02:02:46PM +0300, Vladislav Shpilevoy wrote:
>  function box.schema.space.bless(space)
> -    local index_mt = table.deepcopy(index_mt)
> -    local space_mt = table.deepcopy(space_mt)
> +    -- At first, reference all global index functions. At second,
> +    -- choose an implementation for read operations. They are not
> +    -- in a global index_mt, because they are engine specific.
> +    -- All common index_mt functions must be referenced here to
> +    -- be able to get them using getmetatable(index_object) - if
> +    -- they are only in index_mt, then getmetatable() returns only
> +    -- read ops.
> +    local local_index_mt = table.deepcopy(index_mt)
> +    local_index_mt.__index = function(index, key)
> +        return local_index_mt[key] or index_mt[key]
> +    end

I don't really like that you use the same metatable for all spaces, but
a separate metatable for each index. Besides, because of this one can't
overwrite a built-in method (e.g. 'get') as it will be masked by
local_index_mt. AFAIU you do this, because vinyl and memtx use different
methods (luac vs ffi) for certain operations.

I see the following alternatives:

 1. Leave as is. This isn't all so bad. After all, I doubt anybody's
    actually willing to overwrite built-in methods.

 2. Use two metatables, one for all memtx indexes, another for vinyl,
    and push all changes from the public index_mt to both of these
    metatables. Not quite sure it's technically possible in Lua.

 3. Same as #2, but proscribe modification of built-in methods.

 4. Add a separate engine-specific metatable for the method that differ
    between memtx and vinyl, then define __index as:

    local engine_index_mt = (engine == 'memtx' and
                             memtx_index_mt or vinyl_index_mt)
    index_mt.__index = function(index, key)
        return index_mt[key] or engine_index_mt[key]

Personally, I like #4 most, but it may affect performance.
Anyway, I guess this is up to Kostja to decide which one to choose.

I don't see any other technical problems in the patch, nor in the test.

>      -- true if reading operations may yield
>      local read_yields = space.engine == 'vinyl'
>      local read_ops = {'select', 'get', 'min', 'max', 'count', 'random', 'pairs'}
>      for _, op in ipairs(read_ops) do
>          if read_yields then
>              -- use Lua/C implmenetation
> -            index_mt[op] = index_mt[op .. "_luac"]
> +            local_index_mt[op] = index_mt[op .. "_luac"]
>          else
>              -- use FFI implementation
> -            index_mt[op] = index_mt[op .. "_ffi"]
> +            local_index_mt[op] = index_mt[op .. "_ffi"]
>          end
>      end
>      index_mt.__pairs = index_mt.pairs -- Lua 5.2 compatibility
> @@ -1430,7 +1441,7 @@ function box.schema.space.bless(space)
>      if type(space.index) == 'table' and space.enabled then
>          for j, index in pairs(space.index) do
>              if type(j) == 'number' then
> -                setmetatable(index, index_mt)
> +                setmetatable(index, local_index_mt)
>              end
>          end
>      end