Re: [tarantool-patches] [PATCH v3 1/7] memtx: introduce universal iterator_pool

[Vladimir Davydov <vdavydov.dev@gmail.com>]

On Sun, Feb 24, 2019 at 08:15:04PM +0300, Konstantin Osipov wrote:
> * Vladimir Davydov <vdavydov.dev@gmail.com> [19/02/24 10:01]:
> > On Fri, Feb 22, 2019 at 09:37:25PM +0300, Konstantin Osipov wrote:
> > > * Kirill Shcherbatov <kshcherbatov@tarantool.org> [19/02/22 19:29]:
> > > > Memtx uses separate mempools for iterators of different types.
> > > > Due to the fact that there will be more iterators of different
> > > > sizes in a series of upcoming changes, let's always allocate the
> > > > iterator of the largest size.
> > > 
> > > If rtree iterator is the one which is largest, let's use a
> > > separate pool for it. 
> > > 
> > > In general mempools are rather cheap. Each mempool takes a slab
> > > for ~100 objects and uses no slabs if there are no objects (e.g.
> > > if rtree index is not used, there is no mempool memory for it).
> > 
> > But I'd rather prefer to use the same mempool for all kinds of iterator
> > objects to simplify the code. Take a look at how those mempools are
> > initialized on demand. IMO it looks ugly. Do we really want to save
> > those 500 of bytes that much to put up with that complexity?
> 
> Just like in the recent bps tree performance issue, you don't
> pessimise the code since you never really know how it's going to
> be used.

Oh come on, what pessimization are you talking about in this particular
case? How many iterators can be out there simultaneously? A hundred, a
thousand? 500 bytes overhead per each doesn't seem much, especially
taking into account the fact that you're likely to have a fiber with
16KB stack for each iterator.

Regarding the bps tree performance issue. I see nothing wrong about it.
We've found an issue and we'll surely fix it. There was no point to
think about such a minor optimization until we actually faced the
problem. My point is we should strive to write simple and reliable code
first, and optimize it only if there's a demand, otherwise we risk
turning the code into unmaintainable mess for no good reason.