[tarantool-patches] Re: [PATCH v8 1/3] box: factor fiber_gc out of txn_commit

[Vladislav Shpilevoy <v.shpilevoy@tarantool.org>]

On 31/10/2018 02:08, n.pettik wrote:
> 
>>>> But SQL wants to use some transactional data after commit. It is
>>>> autogenerated identifiers - a list of sequence values generated
>>>> for autoincrement columns and explicit sequence:next() calls.
>>>>
>>>> It is possible to store the list on malloced mem inside Vdbe, but
>>>> it complicates deallocation.
>>> What is the problem with deallocation? AFAIU it is enough to
>>> simply iterate over the list and release each element - not big deal.
>>> If you want to use region, mb it is worth to store separate region
>>> specially for VDBE? We already have it in parser, so what prevents
>>> us for adding the same thing to VDBE? I guess we can store many
>>> things there, not only list of ids. I understand that parser in its turn
>>> has nothing in common (at least it should, except for analyze machinery)
>>> with transaction routines, so separate region is likely to be more
>>> reasonable for parser, but anyway...
>>
>> I've decided to say more details. Parser never yields. This is why we can
>> waste here any resources, rack and ruin everything, but at the end of
>> parsing it should be returned back.
>>
>> Vdbe, on the contrary, yields. So it holds some system resources while
>> other fibers can not use them. If we added a special region to Vdbe, it
>> would steal slabs from the thread's slab cache, while other fibers may
>> want to use it. Hence, when we use one region for all transactional data,
>> including language specific, allocations are much less fragmented over
>> different slabs.
>>
>> Is this explanation decent?
> 
> Quite. I thought that used slabs are marked somehow so that different
> fibers’ regions can’t rely on the same chunk. Probably, I misunderstood
> how internals of our allocation system work. I would better ask you f2f
> someday (or read again Konstantin’s article). Anyway, thanks.
> 
>>
>> Also, I do not agree, that 'deallocation is just iteration and it is
>> ok'. It is O(n) iteration and freeing of heap objects. If a one inserted
>> 10k rows with autogenerated ids, it would waste 10k heap fragments,
>> 10k calls of malloc/free - in my opinion it is an abysmal overhead, but
>> what is more, it can be avoided for free. Instead of 10k free() it boils
>> down to deallocation of N slabs, where N = slab_size / (10k * 8); 8 - size
>> of autogenerated it; slab size is at least 64Kb, so N = 64*1024/80000 <tel:1024/80000>< 1.
>> It takes 1 deallocation vs 10k deallocations. So I think this refactoring
>> is worth.
> 
> Very impressive calculations, however:
> 
> a. I doubt that smb extensively uses queries like
> INSERT INTO t VALUES (NULL, ..), *10k repeats*, (NULL, ..)’
> *Ok, neither I nor you know which queries users execute (or will execute),
>   but anyway your example looks too synthetic.*
> 
> b. Nothing prevents us from counting number of NULLs right in parser
> and allocate memory as single array (one malloc). In this case it would
> be more efficient, I guess, since you don’t need that machinery connected
> with linked list. Btw, why didn’t you consider this variant?

INSERT INTO ... SELECT FROM ...;

Here you can not calculate. Also, it is not possible to calculate
autoids from triggers, box Lua functions. So a list is the only
variant.

>