Re: [tarantool-patches] Re: [PATCH v3 1/1] lua: add key_def lua module

[Alexander Turenko <alexander.turenko@tarantool.org>]

> > +static int
> > +lbox_key_def_compare_with_key(struct lua_State *L)
> > +{
> > +	struct key_def *key_def;
> > +	if (lua_gettop(L) != 3 ||
> > +	    (key_def = luaT_check_key_def(L, 1)) == NULL) {
> > +		return luaL_error(L, "Usage: key_def:"
> > +				     "compare_with_key(tuple, key)");
> > +	}
> > +
> > +	struct tuple *tuple = luaT_key_def_check_tuple(L, key_def, 2);
> > +	if (tuple == NULL)
> > +		return luaT_error(L);
> > +
> > +	size_t key_len;
> > +	const char *key = lbox_encode_tuple_on_gc(L, 3, &key_len);
> > +	uint32_t part_count = mp_decode_array(&key);
> > +	if (key_validate_parts(key_def, key, part_count, true) != 0) {
> > +		tuple_unref(tuple);
> > +		return luaT_error(L);
> > +	}
> > +
> > +	int rc = tuple_compare_with_key(tuple, key, part_count, key_def);
> > +	tuple_unref(tuple);
> > +	lua_pushinteger(L, rc);
> > +	return 1;
> > +}
> 
> This also looks as a terribly inefficient implementation for
> compare.
> 
> Overall, the API looks good to me, while the implementation seems
> to be too inefficient. I would consider changing extract_key() to
> return char *, not struct tuple, the buffer could be allocated on
> transaction region. I also think that compare should not allocate
> memory or create tuples, and it should not call tuple_validate()
> either.
> 
> If this is urgent, I would push since the code quality is very
> good and the api would stable, but I don't see how soo inefficient
> compare functions could be useful.

We can add *_unchecked() functions, which will not perform validation.

I'm against returning char *, because a user can do nothing with it in
Lua.

We can consider a cursor implementation as motivating example. For a
cursor that navigates over an unique index we need :extract_key(), see
[1]. Comparing of tuples is needed if we'll try to implement cursors on
a non-unique index (but maybe better to implement #3898 ('Allow
specifying primary key in iteration by secondary key').

I don't know what Michael F. want when he file #3398 ('lua: introduce
key_def module'). Maybe he just needs merger. Maybe we should introduce
built-in tuple sorter instead of exposing comparators into Lua, which
can be either fast or safe, but not both.

[1]: https://github.com/Totktonada/tarantool-merger-examples/blob/695fc9511685033f4b4b22c0df537a12ddf2eaf6/chunked_example_fast/storage.lua#L97

WBR, Alexander Turenko.