[Tarantool-patches] [PATCH 1/1] sql: introduce UUID field type

[Mergen Imeev]

Hi! Thank you for suggestions. My answers below.

On Wed, May 26, 2021 at 01:58:41AM +0300, Timur Safin wrote:
> There is one major issue I'd like to discuss, please see below...
> 
> : From: Mergen Imeev
> : Subject: Re: [Tarantool-patches] [PATCH 1/1] sql: introduce UUID field type
> : 
> ...
> 
> : +static inline int
> : +str_to_uuid(struct Mem *mem)
> : +{
> : +	assert(mem->type == MEM_TYPE_STR);
> : +	struct tt_uuid uuid;
> : +	if (mem->n != UUID_STR_LEN ||
> : +	    tt_uuid_from_string(tt_cstr(mem->z, mem->n), &uuid) != 0)
> 
> Here lies the limitation I've specifically wanted to avoid and put
> below to RFC:
> 
> "The most complete implementation seemingly is in PostgreSQL, 
> which allows various relaxed formats for string literals 
> which may be accepted as UUID - the plan is to be as close
> as possible to PostgreSQL here."
> 
> Checking for UUID_STR_LEN restricts only to 1 specific form with
> dashes in specific positions, but we would rather want to allow
> relaxed formats also, with some dashes omitted. Also with optional
> braces around. i.e. putting this as regexp
> 
>   \{?[0-9a-f]{8}-?[0-9a-f]{4}-?[0-9a-f]{4}-?[0-9a-f]{4}-?[0-9a-f]{12}\}?
> 
> And yes, I mean we need to reimplement tt_uuid_from_string() differently.
> (If you were asking me then I'd use RE2C for that purpose, but that up
> to you)
> 
Addition of support of other versions of UUID should not be done as part of this
issue, since it is problem of BOX, not just SQL.

> : +		return -1;
> : +	mem_set_uuid(mem, &uuid);
> : +	return 0;
> : +}
> : +
> :  static inline int
> :  str_to_bool(struct Mem *mem)
> :  {
> ...
> : diff --git a/src/box/sql/mem.h b/src/box/sql/mem.h
> : index 15d97da0e..b3cd5c545 100644
> : --- a/src/box/sql/mem.h
> : +++ b/src/box/sql/mem.h
> : @@ -30,6 +30,7 @@
> :   * SUCH DAMAGE.
> :   */
> :  #include "box/field_def.h"
> : +#include "uuid/tt_uuid.h"
> : 
> :  struct sql;
> :  struct Vdbe;
> : @@ -47,10 +48,11 @@ enum mem_type {
> :  	MEM_TYPE_MAP		= 1 << 6,
> :  	MEM_TYPE_BOOL		= 1 << 7,
> :  	MEM_TYPE_DOUBLE		= 1 << 8,
> : -	MEM_TYPE_INVALID	= 1 << 9,
> : -	MEM_TYPE_FRAME		= 1 << 10,
> : -	MEM_TYPE_PTR		= 1 << 11,
> : -	MEM_TYPE_AGG		= 1 << 12,
> : +	MEM_TYPE_UUID		= 1 << 9,
> : +	MEM_TYPE_INVALID	= 1 << 10,
> : +	MEM_TYPE_FRAME		= 1 << 11,
> : +	MEM_TYPE_PTR		= 1 << 12,
> : +	MEM_TYPE_AGG		= 1 << 13,
> 
> I guess there should be no incompatibility problems here, but just
> in case I ask - why did you insert new constant to the middle of prior list?
> 
Before MEM_TYPE_INVALID are types that are accesisble to user. Starting with
MEM_TYPE_INVALID are for internal use.

> Also I see there is another omission from RFC - there is nothing added 
> for SQL built-in support to generate UUID from within SQL mode.
> 
>   "Introduce UUID([version#]) functions which would allow generating
>    any particular type of GUID. If version argument is omitted then
>    we generate UUID v4 (as used by default in box);"
> 
I will add this function.

> Regards,
> Timur
>