Hi Igor,
What Ignite version do you use? Recent versions do not read real rows for
the simplest "SELECT COUNT(*) FROM table" query. Instead it goes over index
and simply count records, so no copying happens. But let me express again -
this is not the case in general. If any expression appears, we have to read
the row from offheap. In this case the whole key-value pair is read from
offheap including non-SQL fields. This is so because we do not know in
advance what parts of key or value will be needed. Even if specific field
is not marked as @QuerySqlField, it may be needed if the whole key or value
is requested ("SELECT _KEY FROM table"), or if it is used in some custom
function, etc..
In future versions we may re-approach this, and try to read only fields
required for query from offheap. But this will require significant changes
to both storage layer and query engine, so this is not an easy task.
On Thu, Oct 18, 2018 at 2:42 AM Igor Tanackovic <igor.tanacko...@gmail.com>
wrote:
> Dmitriy,
>
> Correct me if I’m wrong, but the concept is to store everything off heap -
> which is perfectly fine :). So, the question is how SQL engine actually
> works. Analyzing profiler object generation and sizes on heap, I’ve learned
> that for each query entire rows in question are copied on heap which, I’d
> say is a little bit inefficient (more or less depending on domain object).
> Maybe an example will help to further elaborate on the issue.
>
> Lets say we have an object ORGANIZATION which has only two query enabled
> fields (annotated with @QuerySqlField) and 20 more value fields
> (text/blobs, …) which are not used in queries (not annotated with
> @QuerySqlField and thus not visible by SQL engine). So, how COUNT works?
> Will SQL engine copy on heap just a part of ORGANIZATION rows (2 fields
> that might be part of COUNT query) or entire rows (sql enabled fields +
> text/blob fields…)?
>
> From what I saw during profiling, I’d say entire rows will be copied on
> heap. Just to compare, this is our, real world example:
>
> Object A (~300k rows, each row has 5 columns, all sql enabled), COUNT(*)
> takes ~2MB on heap
> Object B (~25k rows, each row has 16 columns, 5 sql enabled fields and 11
> text value fields), COUNT(*) takes ~60MB on heap - nearly 30 times more
>
>
> Regards,
> Igor
>
>
>
> > On Oct 18, 2018, at 12:24 AM, Dmitriy Setrakyan <dsetrak...@apache.org>
> wrote:
> >
> > I do not understand - why are we copying values on-heap? If we must copy
> > something, why not copy rows into some off-heap space and read what we
> have
> > from there? At least this way we will not burden the JVM with extra GC
> > pressure.
> >
> > D.
> >
> > On Wed, Oct 17, 2018 at 10:33 AM Igor Tanackovic <
> igor.tanacko...@gmail.com>
> > wrote:
> >
> >> Vladimir,
> >>
> >> Thanks for explanation... that’s true - ignite never deserialize values
> >> during query execution. The point here is why copying fields to heap
> that
> >> sql engine could not use (not annotated as QuerySqlField)?
> >> SQL count is a perfect example where you can benefit (heap space and gc)
> >> copying only field sql engine could use.
> >>
> >> Regards,
> >> Igor
> >>
> >> On Wed, Oct 17, 2018 at 18:47 Vladimir Ozerov <voze...@gridgain.com>
> >> wrote:
> >>
> >>> Hi Igor,
> >>>
> >>> We never deserialize values during query execution. Instead, we copy
> the
> >>> row to heap and extract fields as needed. In general case it is
> >> impossible
> >>> to avoid reading the whole row because we do not know whether this is
> >>> COUNT(*) or COUNT(*) WHERE <something> or COUNT(<something>) WHERE
> >>> <something>.
> >>> We do handle plain COUNT(*) as speical case and iterate over index
> only,
> >>> but event this simple query cannot avoid row reads in general case when
> >> new
> >>> snapshot mode is enabled, because an entry in the index may be not
> >> visible
> >>> to current transaction.
> >>>
> >>> On Wed, Oct 17, 2018 at 7:19 PM igor.tanackovic <
> >> igor.tanacko...@gmail.com
> >>>>
> >>> wrote:
> >>>
> >>>> Hello,
> >>>>
> >>>> Seems that SQL engine always deserialize whole objects instead of
> using
> >>>> just
> >>>> SQL enabled fields (annotated with @QuerySqlField). This may have a
> >> huge
> >>>> impact on Ignite heap usage and GC overhead as well.
> >>>>
> >>>> For example, we have a cache holding big objects but with only two sql
> >>>> query
> >>>> fields which for each query execution (SELECT COUNT(*) FROM 'cache')
> >>>> consumes large amount on heap memory (~300MB). As a proof of concept,
> >> we
> >>>> divided the same cache to *index* cache with only sql query field and
> a
> >>>> *data* holding whole object for materialization. The same query
> (SELECT
> >>>> COUNT(*) FROM 'index-cache') consumes ~25 time less memory! The same
> is
> >>>> true
> >>>> for all other queries.
> >>>>
> >>>> The obvious workaround would be to always have separated regions for
> >>>> indexes
> >>>> (sql query enabled region) and a data/value region for
> materialization,
> >>> but
> >>>> it might be a good idea to fix this in a systematic way during off
> heap
> >>>> deserialization.
> >>>>
> >>>> Regards,
> >>>> Igor
> >>>>
> >>>>
> >>>>
> >>>> --
> >>>> Sent from: http://apache-ignite-developers.2346864.n4.nabble.com/
> >>>>
> >>>
> >>
>
>
