Re: Some questions on intermediate serialization in Pig

[Gianmarco De Francisci Morales <gd...@apache.org>]

So, to recap.

InterSedes writes the R1/R2/R3 thing.
I am quite sure it is done for splittability purposes.
The RawComparators, as well as InterStorage, operate on binary data that
does not see this thing.

DISCLAIMER: Guesswork below!

My wild guess is that InterRecordReader has nothing to do with the
RawComparator.
The former is used at the input of the map phase (and InterRecordWriter is
used at the output of the reduce phase) while RawComparator is used at the
boundary between map and reduce phase.
If you look at PigGenericMapReduce you will see that the mapper/reducer
always writes a PigNullableWritable as a key. I think this is the place
where the RawComparator actually plays a role, which means it does not
directly see the R1/R2/R3 because they are stripped out by the RecordReader.


Cheers,

--
Gianmarco



On Sun, May 27, 2012 at 10:28 AM, Jonathan Coveney <jc...@gmail.com>wrote:

> Ashutosh, that definitely does help. Thanks for lending your insight. I
> think the thing I have little color on at the moment is the relationship
> between those raw bits ie RECORD_1 RECORD_2 RECORD_3 TUPLE_BITS and so on,
> and then the various byte[] compare functions.
>
> 2012/5/27 Ashutosh Chauhan <ha...@apache.org>
>
> > Hey Jon,
> >
> > You raised some interesting question. I don't have answer for all, but
> have
> > for few.
> >
> > * BinStorage is a legacy format which was used for intermediate
> > serialization between MR jobs earlier. It is no longer used but is there
> > because unfortunately folks have stored their end-data using BinStorage,
> > even though it was considered internal format and subject to change. The
> > reason folks chose to store data using it was BinStorage was schema
> aware,
> > so once u wrote end-data with it, you can reload it without specifying
> > schema. This feature led to its (mis)use. See
> > https://issues.apache.org/jira/browse/PIG-798 for some related bugs
> around
> > this.
> >
> > * I think you have a correct intuition that in addition to identify tuple
> > boundaries, R1,R2,R3 is also used to identify block boundaries, that is
> to
> > make file splittable. Since, then you can arbitrarily split the files
> among
> > multiple mappers and they will know where does their first record starts.
> >
> > Hope it helps,
> > Ashutosh
> >
> > On Sat, May 26, 2012 at 9:04 PM, Jonathan Coveney <jcoveney@gmail.com
> > >wrote:
> >
> > > I appreciate it, Gianmarco :)
> > >
> > > 2012/5/26 Gianmarco De Francisci Morales <gd...@apache.org>
> > >
> > > > I am not sure, but I will have a look at it (I implemented the raw
> > > > comparator for secondary sort).
> > > > I don't remember having to deal with this issue.
> > > >
> > > > Cheers,
> > > > --
> > > > Gianmarco
> > > >
> > > >
> > > >
> > > >
> > > > On Fri, May 25, 2012 at 11:07 PM, Jonathan Coveney <
> jcoveney@gmail.com
> > > > >wrote:
> > > >
> > > > > I'll just bump this once. The main thing I'm still unsure on is
> just
> > > the
> > > > > relationship various raw comparators, Pig, and hadoop. If we're
> > > > serializing
> > > > > RECORD_1, RECORD_2, RECORD_3, Tuple, RECORD_1, RECORD_2, RECORD_3,
> > > > > Tuple, RECORD_1, RECORD_2, RECORD_3, Tuple, RECORD_1, RECORD_2,
> > > RECORD_3,
> > > > > Tuple, RECORD_1, RECORD_2, RECORD_3, Tuple, and so on, how come it
> > > > appears
> > > > > that the raw comparators aren't aware of it?
> > > > >
> > > > > 2012/5/23 Jonathan Coveney <jc...@gmail.com>
> > > > >
> > > > > > And one more question to pile on:
> > > > > >
> > > > > > What defines the binary data that the raw tuple comparator will
> be
> > > run
> > > > > on?
> > > > > > It seems like that it comes from hadoop, and the format generally
> > > makes
> > > > > > sense (you get bytes and do with them what you will). The thing
> > that
> > > > > > confuses me is why don't you have to deal with the
> > > > > > RECORD_1/RECORD_2/RECORD_3/etc hooplah? Interstorage deals with
> all
> > > of
> > > > > that
> > > > > > and reads a deserialized tuple...so at what point do you get
> binary
> > > > Tuple
> > > > > > data that doesn't have all of the split stuff? I'll keep digging
> > > > through
> > > > > > but this is where my ignorance of the technicalities of the MR
> > layer
> > > > > comes
> > > > > > in...
> > > > > >
> > > > > > 2012/5/23 Jonathan Coveney <jc...@gmail.com>
> > > > > >
> > > > > >> Another question is clarifying what BinStorage does compared to
> > > > > >> InterStorage. It looks like it might just be a legacy storage
> > > format?
> > > > > >>
> > > > > >> I'm assuming that you do the R_1/R_2/R_3 to be able to find the
> > next
> > > > > >> Tuple in the stream, but once you do that, can't you just read a
> > > > tuple,
> > > > > and
> > > > > >> then read skip 12 bytes (3 ints), and keep reading?
> > > > > >>
> > > > > >>
> > > > > >> 2012/5/23 Jonathan Coveney <jc...@gmail.com>
> > > > > >>
> > > > > >>> I'm trying to understand how intermediate serialization in Pig
> > > works
> > > > at
> > > > > >>> a deeper level (understanding the whole code path, not just
> > > > > BinInterSedes
> > > > > >>> in its own vaccuum). Right now I am looking at
> > > > > >>> InterRecordReader/InterRecordWriter/InterStorage. Is that the
> > right
> > > > > place
> > > > > >>> to look for understanding how BinInterSedes is actually called?
> > > > > >>>
> > > > > >>> Further, I'm trying to better understanding the
> > > > > >>> RECORD_1/RECORD_2/RECORD_3 thing. My guess is that it's to make
> > the
> > > > > file
> > > > > >>> splittable? But I'm not really sure. I'd love any pointers
> about
> > > > where
> > > > > to
> > > > > >>> look for how BinInterSedes is used, and how intermediate
> storage
> > > > > happens.
> > > > > >>>
> > > > > >>> Thanks!
> > > > > >>> Jon
> > > > > >>>
> > > > > >>
> > > > > >>
> > > > > >
> > > > >
> > > >
> > >
> >
>

Re: Some questions on intermediate serialization in Pig

[Jonathan Coveney <jc...@gmail.com>]

Gianmarco, I think you're absolutely correct. Thanks for weighing in :)

2012/6/9 Gianmarco De Francisci Morales <gd...@apache.org>

> So, to recap.
>
> InterSedes writes the R1/R2/R3 thing.
> I am quite sure it is done for splittability purposes.
> The RawComparators, as well as InterStorage, operate on binary data that
> does not see this thing.
>
> DISCLAIMER: Guesswork below!
>
> My wild guess is that InterRecordReader has nothing to do with the
> RawComparator.
> The former is used at the input of the map phase (and InterRecordWriter is
> used at the output of the reduce phase) while RawComparator is used at the
> boundary between map and reduce phase.
> If you look at PigGenericMapReduce you will see that the mapper/reducer
> always writes a PigNullableWritable as a key. I think this is the place
> where the RawComparator actually plays a role, which means it does not
> directly see the R1/R2/R3 because they are stripped out by the
> RecordReader.
>
>
> Cheers,
>
> --
> Gianmarco
>
>
>
> On Sun, May 27, 2012 at 10:28 AM, Jonathan Coveney <jcoveney@gmail.com
> >wrote:
>
> > Ashutosh, that definitely does help. Thanks for lending your insight. I
> > think the thing I have little color on at the moment is the relationship
> > between those raw bits ie RECORD_1 RECORD_2 RECORD_3 TUPLE_BITS and so
> on,
> > and then the various byte[] compare functions.
> >
> > 2012/5/27 Ashutosh Chauhan <ha...@apache.org>
> >
> > > Hey Jon,
> > >
> > > You raised some interesting question. I don't have answer for all, but
> > have
> > > for few.
> > >
> > > * BinStorage is a legacy format which was used for intermediate
> > > serialization between MR jobs earlier. It is no longer used but is
> there
> > > because unfortunately folks have stored their end-data using
> BinStorage,
> > > even though it was considered internal format and subject to change.
> The
> > > reason folks chose to store data using it was BinStorage was schema
> > aware,
> > > so once u wrote end-data with it, you can reload it without specifying
> > > schema. This feature led to its (mis)use. See
> > > https://issues.apache.org/jira/browse/PIG-798 for some related bugs
> > around
> > > this.
> > >
> > > * I think you have a correct intuition that in addition to identify
> tuple
> > > boundaries, R1,R2,R3 is also used to identify block boundaries, that is
> > to
> > > make file splittable. Since, then you can arbitrarily split the files
> > among
> > > multiple mappers and they will know where does their first record
> starts.
> > >
> > > Hope it helps,
> > > Ashutosh
> > >
> > > On Sat, May 26, 2012 at 9:04 PM, Jonathan Coveney <jcoveney@gmail.com
> > > >wrote:
> > >
> > > > I appreciate it, Gianmarco :)
> > > >
> > > > 2012/5/26 Gianmarco De Francisci Morales <gd...@apache.org>
> > > >
> > > > > I am not sure, but I will have a look at it (I implemented the raw
> > > > > comparator for secondary sort).
> > > > > I don't remember having to deal with this issue.
> > > > >
> > > > > Cheers,
> > > > > --
> > > > > Gianmarco
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > On Fri, May 25, 2012 at 11:07 PM, Jonathan Coveney <
> > jcoveney@gmail.com
> > > > > >wrote:
> > > > >
> > > > > > I'll just bump this once. The main thing I'm still unsure on is
> > just
> > > > the
> > > > > > relationship various raw comparators, Pig, and hadoop. If we're
> > > > > serializing
> > > > > > RECORD_1, RECORD_2, RECORD_3, Tuple, RECORD_1, RECORD_2,
> RECORD_3,
> > > > > > Tuple, RECORD_1, RECORD_2, RECORD_3, Tuple, RECORD_1, RECORD_2,
> > > > RECORD_3,
> > > > > > Tuple, RECORD_1, RECORD_2, RECORD_3, Tuple, and so on, how come
> it
> > > > > appears
> > > > > > that the raw comparators aren't aware of it?
> > > > > >
> > > > > > 2012/5/23 Jonathan Coveney <jc...@gmail.com>
> > > > > >
> > > > > > > And one more question to pile on:
> > > > > > >
> > > > > > > What defines the binary data that the raw tuple comparator will
> > be
> > > > run
> > > > > > on?
> > > > > > > It seems like that it comes from hadoop, and the format
> generally
> > > > makes
> > > > > > > sense (you get bytes and do with them what you will). The thing
> > > that
> > > > > > > confuses me is why don't you have to deal with the
> > > > > > > RECORD_1/RECORD_2/RECORD_3/etc hooplah? Interstorage deals with
> > all
> > > > of
> > > > > > that
> > > > > > > and reads a deserialized tuple...so at what point do you get
> > binary
> > > > > Tuple
> > > > > > > data that doesn't have all of the split stuff? I'll keep
> digging
> > > > > through
> > > > > > > but this is where my ignorance of the technicalities of the MR
> > > layer
> > > > > > comes
> > > > > > > in...
> > > > > > >
> > > > > > > 2012/5/23 Jonathan Coveney <jc...@gmail.com>
> > > > > > >
> > > > > > >> Another question is clarifying what BinStorage does compared
> to
> > > > > > >> InterStorage. It looks like it might just be a legacy storage
> > > > format?
> > > > > > >>
> > > > > > >> I'm assuming that you do the R_1/R_2/R_3 to be able to find
> the
> > > next
> > > > > > >> Tuple in the stream, but once you do that, can't you just
> read a
> > > > > tuple,
> > > > > > and
> > > > > > >> then read skip 12 bytes (3 ints), and keep reading?
> > > > > > >>
> > > > > > >>
> > > > > > >> 2012/5/23 Jonathan Coveney <jc...@gmail.com>
> > > > > > >>
> > > > > > >>> I'm trying to understand how intermediate serialization in
> Pig
> > > > works
> > > > > at
> > > > > > >>> a deeper level (understanding the whole code path, not just
> > > > > > BinInterSedes
> > > > > > >>> in its own vaccuum). Right now I am looking at
> > > > > > >>> InterRecordReader/InterRecordWriter/InterStorage. Is that the
> > > right
> > > > > > place
> > > > > > >>> to look for understanding how BinInterSedes is actually
> called?
> > > > > > >>>
> > > > > > >>> Further, I'm trying to better understanding the
> > > > > > >>> RECORD_1/RECORD_2/RECORD_3 thing. My guess is that it's to
> make
> > > the
> > > > > > file
> > > > > > >>> splittable? But I'm not really sure. I'd love any pointers
> > about
> > > > > where
> > > > > > to
> > > > > > >>> look for how BinInterSedes is used, and how intermediate
> > storage
> > > > > > happens.
> > > > > > >>>
> > > > > > >>> Thanks!
> > > > > > >>> Jon
> > > > > > >>>
> > > > > > >>
> > > > > > >>
> > > > > > >
> > > > > >
> > > > >
> > > >
> > >
> >
>