[#HADOOP-2085] Map-side joins on sorted, equally-partitioned datasets

Hadoop Common

Map-side joins on sorted, equally-partitioned datasets

Created: 22/Oct/07 03:59 AM Updated: 08/Jul/09 04:52 PM

Return to search

Component/s:

None

Affects Version/s:

None

Fix Version/s:

0.16.0

Time Tracking:

Not Specified

File Attachments:

File Name	Date Attached	Attached By	Size
2085-2.patch	2007-10-26 07:21 PM	Chris Douglas	101 kB
2085-3.patch	2007-10-27 11:37 PM	Chris Douglas	109 kB
2085-4.patch	2007-11-09 01:34 AM	Chris Douglas	108 kB
2085-5.patch	2007-11-16 09:42 PM	Chris Douglas	112 kB
2085.patch	2007-10-22 04:00 AM	Chris Douglas	101 kB

Resolution Date:

07/Dec/07 10:06 PM

Description

« Hide

Motivation

Given a set of sorted datasets keyed with the same class and yielding equal
partitions, it is possible to effect a join of those datasets prior to the
map. This could save costs in re-partitioning, sorting, shuffling, and
writing out data required in the general case.

Interface

The attached code offers the following interface to users of these classes.

property	required	value
mapred.join.expr	yes	Join expression to effect over input data
mapred.join.keycomparator	no	`WritableComparator` class to use for comparing keys
mapred.join.define.<ident>	no	Class mapped to identifier in join expression

The join expression understands the following grammar:

func ::= <ident>([<func>,]*<func>)
func ::= tbl(<class>,"<path>");

Operations included in this patch are partitioned into one of two types:
join operations emitting tuples and "multi-filter" operations emitting a
single value from (but not necessarily included in) a set of input values.
For a given key, each operation will consider the cross product of all
values for all sources at that node.

Identifiers supported by default:

identifier	type	description
inner	Join	Full inner join
outer	Join	Full outer join
override	MultiFilter	For a given key, prefer values from the rightmost source

A user of this class must set the InputFormat for the job to
CompositeInputFormat and define a join expression accepted by the preceding
grammar. For example, both of the following are acceptable:

inner(tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
          "hdfs://host:8020/foo/bar"),
      tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
          "hdfs://host:8020/foo/baz"))

outer(override(tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
                   "hdfs://host:8020/foo/bar"),
               tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
                   "hdfs://host:8020/foo/baz")),
      tbl(org.apache.hadoop.mapred/SequenceFileInputFormat.class,
          "hdfs://host:8020/foo/rab"))

CompositeInputFormat includes a handful of convenience methods to aid
construction of these verbose statements.

As in the second example, joins may be nested. Users may provide a
comparator class in the mapred.join.keycomparator property to
specify the ordering of their keys, or accept the default comparator as
returned by WritableComparator.get(keyclass).

Users can specify their own join operations, typically by overriding
JoinRecordReader or MultiFilterRecordReader and mapping that class
to an identifier in the join expression using the
mapred.join.define.ident property, where ident is the identifier
appearing in the join expression. Users may elect to emit- or modify- values
passing through their join operation. Consulting the existing operations for
guidance is recommended. Adding arguments is considerably more complex (and
only partially supported), as one must also add a Node type to the parse
tree. One is probably better off extending RecordReader in most cases.

Design

As alluded to above, the design defines inner (Composite) and leaf (Wrapped)
types for the join tree. Delegation satisfies most requirements of the
InputFormat contract, particularly validateInput and getSplits.
Most of the work in this patch concerns getRecordReader. The
CompositeInputFormat itself delegates to the parse tree generated by
Parser.

Hierarchical Joins

Each RecordReader from the user must be "wrapped", since effecting a
join requires the framework to track the head value from each source. Since
the cross product of all values for each composite level of the join is
emitted to its parent, all sources ¹ must be capable of repeating the
values for the current key. To avoid keeping an excessive number of copies
(one per source per level), each composite requests its children to populate
a JoinCollector with an iterator over its values. This way, there is
only one copy of the current key for each composite node, the head key-value
pair for each leaf, and storage at each leaf for all the values matching the
current key at the parent collector (if it is currently participating in a
join at the root). Strategies have been employed to avoid excessive copying
when filling a user-provided Writable, but they have been conservative
(e.g. in MultiFilterRecordReader, the value emitted is cloned in case
the user modifies the value returned, possibly changing the state of a
JoinCollector in the tree). For example, if the following sources
contain these key streams:

A: 0  0   1    1     2        ...
B: 1  1   1    1     2        ...
C: 1  6   21   107   ...
D: 6  28  496  8128  33550336 ...

Let A-D be wrapped sources and x,y be composite operations. If the
expression is of the form x(A, y(B,C,D)), then when the current key at
the root is 1 the tree may look like this:

            x (1, [ I(A), [ I(y) ] ] )
          /   \
         W     y (1, [ I(B), I(C), EMPTY ])
         |   / | \
         |  W  W  W
         |  |  |  D (6, V~6~) => EMPTY
         |  |  C (6, V~6~)    => V~1.1~ @1.1
         |  B (2, V~2~)       => V~1,1~ V~1,2~ V~1,3~ V~1,4~ @1,3
         A (2, V~2~)          => V~1,1~ V~1,2~ @1,2

A JoinCollector from x will have been created by requesting an
iterator from A and another from y. The iterator at y is built by
requesting iterators from B, C, and D. Since D doesn't contain the
key 1, it returns an empty iterator. Since the value to return for a given
join is a Writable provided by the user, the iterators returned are also
responsible for writing the next value in that stream. For multilevel joins
passing through a subclass of JoinRecordReader, the value produced will
contain tuples within tuples; iterators for composites delegate to
sub-iterators responsible for filling the value in the tuple at the position
matching their position in the composite. In a sense, the only iterators
that write to a tuple are the RecordReader s at the leaves. Note that
this also implies that emitted tuples may not contain values from each
source, but they will always have the same capacity.

Writables

Writable objects- including InputSplit s and TupleWritable s-
encode themselves in the following format:

<count><class1><class2>...<classn><obj1><obj2>...<objn>

The inefficiency is regrettable- particularly since this overhead is
incurred for every instance and most often the tuples emitted will be
processed only within the map- but the encoding satisfies the Writable
contract well enough to be emitted to the reducer, written to disk, etc. It
is hoped that general compression will trim the most egregious waste. It
should be noted that the framework does not actually write out a tuple (i.e.
does not suffer for this deficiency) unless emitting one from
MultiFilterRecordReader (a rare case in practice, it is hoped).

Extensibility

Adding arguments to expressions is more difficult. One would need to include
a Node type for the parser, which requires some knowledge of its inner
workings. The model in this area is crude and requires refinement before it
can be "extensible" by a reasonable definition.

Performance

I have no numbers.

Notes

1. This isn't strictly true. The "leftmost" source will never need to repeat
itself. Adding a pseudo-ResettableIterator to handle this case would be
a welcome addition.

2. Note that- even if reset- the override will only loop through the values
in the rightmost key, instead of repeating that series a number of times
equal to the cardinality of the cross product of the discarded streams
(regrettably, looking at the code of OverrideRecordReader is more
illustrative than this explanation).

Description

Motivation

Given a set of sorted datasets keyed with the same class and yielding equal partitions, it is possible to effect a join of those datasets prior to the map. This could save costs in re-partitioning, sorting, shuffling, and writing out data required in the general case.

Interface

The attached code offers the following interface to users of these classes.

property	required	value
mapred.join.expr	yes	Join expression to effect over input data
mapred.join.keycomparator	no	`WritableComparator` class to use for comparing keys
mapred.join.define.<ident>	no	Class mapped to identifier in join expression

The join expression understands the following grammar:

func ::= <ident>([<func>,]*<func>)
func ::= tbl(<class>,"<path>");

Operations included in this patch are partitioned into one of two types: join operations emitting tuples and "multi-filter" operations emitting a single value from (but not necessarily included in) a set of input values. For a given key, each operation will consider the cross product of all values for all sources at that node. Identifiers supported by default:

identifier	type	description
inner	Join	Full inner join
outer	Join	Full outer join
override	MultiFilter	For a given key, prefer values from the rightmost source

A user of this class must set the InputFormat for the job to CompositeInputFormat and define a join expression accepted by the preceding grammar. For example, both of the following are acceptable:

inner(tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
          "hdfs://host:8020/foo/bar"),
      tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
          "hdfs://host:8020/foo/baz"))

outer(override(tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
                   "hdfs://host:8020/foo/bar"),
               tbl(org.apache.hadoop.mapred.SequenceFileInputFormat.class,
                   "hdfs://host:8020/foo/baz")),
      tbl(org.apache.hadoop.mapred/SequenceFileInputFormat.class,
          "hdfs://host:8020/foo/rab"))

CompositeInputFormat includes a handful of convenience methods to aid construction of these verbose statements. As in the second example, joins may be nested. Users may provide a comparator class in the mapred.join.keycomparator property to specify the ordering of their keys, or accept the default comparator as returned by WritableComparator.get(keyclass). Users can specify their own join operations, typically by overriding JoinRecordReader or MultiFilterRecordReader and mapping that class to an identifier in the join expression using the mapred.join.define.ident property, where ident is the identifier appearing in the join expression. Users may elect to emit- or modify- values passing through their join operation. Consulting the existing operations for guidance is recommended. Adding arguments is considerably more complex (and only partially supported), as one must also add a Node type to the parse tree. One is probably better off extending RecordReader in most cases.

Design

As alluded to above, the design defines inner (Composite) and leaf (Wrapped) types for the join tree. Delegation satisfies most requirements of the InputFormat contract, particularly validateInput and getSplits. Most of the work in this patch concerns getRecordReader. The CompositeInputFormat itself delegates to the parse tree generated by Parser.

Hierarchical Joins

Each RecordReader from the user must be "wrapped", since effecting a join requires the framework to track the head value from each source. Since the cross product of all values for each composite level of the join is emitted to its parent, all sources ¹ must be capable of repeating the values for the current key. To avoid keeping an excessive number of copies (one per source per level), each composite requests its children to populate a JoinCollector with an iterator over its values. This way, there is only one copy of the current key for each composite node, the head key-value pair for each leaf, and storage at each leaf for all the values matching the current key at the parent collector (if it is currently participating in a join at the root). Strategies have been employed to avoid excessive copying when filling a user-provided Writable, but they have been conservative (e.g. in MultiFilterRecordReader, the value emitted is cloned in case the user modifies the value returned, possibly changing the state of a JoinCollector in the tree). For example, if the following sources contain these key streams:

A: 0  0   1    1     2        ...
B: 1  1   1    1     2        ...
C: 1  6   21   107   ...
D: 6  28  496  8128  33550336 ...

Let A-D be wrapped sources and x,y be composite operations. If the expression is of the form x(A, y(B,C,D)), then when the current key at the root is 1 the tree may look like this:

            x (1, [ I(A), [ I(y) ] ] )
          /   \
         W     y (1, [ I(B), I(C), EMPTY ])
         |   / | \
         |  W  W  W
         |  |  |  D (6, V~6~) => EMPTY
         |  |  C (6, V~6~)    => V~1.1~ @1.1
         |  B (2, V~2~)       => V~1,1~ V~1,2~ V~1,3~ V~1,4~ @1,3
         A (2, V~2~)          => V~1,1~ V~1,2~ @1,2

A JoinCollector from x will have been created by requesting an iterator from A and another from y. The iterator at y is built by requesting iterators from B, C, and D. Since D doesn't contain the key 1, it returns an empty iterator. Since the value to return for a given join is a Writable provided by the user, the iterators returned are also responsible for writing the next value in that stream. For multilevel joins passing through a subclass of JoinRecordReader, the value produced will contain tuples within tuples; iterators for composites delegate to sub-iterators responsible for filling the value in the tuple at the position matching their position in the composite. In a sense, the only iterators that write to a tuple are the RecordReader s at the leaves. Note that this also implies that emitted tuples may not contain values from each source, but they will always have the same capacity.

Writables

Writable objects- including InputSplit s and TupleWritable s- encode themselves in the following format:

<count><class1><class2>...<classn><obj1><obj2>...<objn>

The inefficiency is regrettable- particularly since this overhead is incurred for every instance and most often the tuples emitted will be processed only within the map- but the encoding satisfies the Writable contract well enough to be emitted to the reducer, written to disk, etc. It is hoped that general compression will trim the most egregious waste. It should be noted that the framework does not actually write out a tuple (i.e. does not suffer for this deficiency) unless emitting one from MultiFilterRecordReader (a rare case in practice, it is hoped).

Extensibility

The join framework is modestly extensible. Practically, users seeking to add their own identifiers to join expressions are limited to extending JoinRecordReader and MultiFilterRecordReader. There is considerable latitude within these constraints, as illustrated in OverrideRecordReader, where values in child RecordReader s are skipped instead of incurring the overhead of building the iterator (that will inevitably be discarded).² For most cases, the user need only implement the combine and/or emit methods in their subclass. It is expected that most will find that the three default operations will suffice. Adding arguments to expressions is more difficult. One would need to include a Node type for the parser, which requires some knowledge of its inner workings. The model in this area is crude and requires refinement before it can be "extensible" by a reasonable definition.

Performance

I have no numbers. Notes 1. This isn't strictly true. The "leftmost" source will never need to repeat itself. Adding a pseudo-ResettableIterator to handle this case would be a welcome addition. 2. Note that- even if reset- the override will only loop through the values in the rightmost key, instead of repeating that series a number of times equal to the cardinality of the cross product of the discarded streams (regrettably, looking at the code of OverrideRecordReader is more illustrative than this explanation).

Show »

All

Comments

Work Log

Change History

Subversion Commits

Sort Order:

[ Permlink | « Hide ]

Tsz Wo (Nicholas), SZE added a comment - 24/Oct/07 01:57 AM

By assuming certain properties of the input datasets, the join operation might be performed more efficiently. I think this is an interesting observation.

I tried to read your patch but still cannot fully understand it. It would be great if you can give an example (like WordCount) to show how to use the new codes.

[ Permlink | « Hide ]

Hadoop QA added a comment - 26/Oct/07 08:14 AM

-1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12368110/2085.patch
against trunk revision r588341.

@author +1. The patch does not contain any @author tags.

javadoc -1. The javadoc tool appears to have generated messages.

javac +1. The applied patch does not generate any new compiler warnings.

findbugs -1. The patch appears to introduce 4 new Findbugs warnings.

core tests +1. The patch passed core unit tests.

contrib tests -1. The patch failed contrib unit tests.

Test results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1003/testReport/
Findbugs warnings: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1003/artifact/trunk/build/test/findbugs/newPatchFindbugsWarnings.html
Checkstyle results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1003/artifact/trunk/build/test/checkstyle-errors.html
Console output: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1003/console

This message is automatically generated.

[ Permlink | « Hide ]

Chris Douglas added a comment - 26/Oct/07 07:21 PM

Fixed findbugs warnings, addressed javadoc, changed Token type to accommodate Nicholas's feedback.

[ Permlink | « Hide ]

Hadoop QA added a comment - 26/Oct/07 10:22 PM

-1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12368513/2085-2.patch
against trunk revision r588778.

@author +1. The patch does not contain any @author tags.

javadoc +1. The javadoc tool did not generate any warning messages.

javac +1. The applied patch does not generate any new compiler warnings.

findbugs -1. The patch appears to introduce 3 new Findbugs warnings.

core tests -1. The patch failed core unit tests.

contrib tests +1. The patch passed contrib unit tests.

Test results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1012/testReport/
Findbugs warnings: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1012/artifact/trunk/build/test/findbugs/newPatchFindbugsWarnings.html
Checkstyle results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1012/artifact/trunk/build/test/checkstyle-errors.html
Console output: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1012/console

This message is automatically generated.

[ Permlink | « Hide ]

Chris Douglas added a comment - 27/Oct/07 11:37 PM

More findbugs, added an example

[ Permlink | « Hide ]

Hadoop QA added a comment - 29/Oct/07 05:56 AM

-1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12368550/2085-3.patch
against trunk revision r588778.

@author +1. The patch does not contain any @author tags.

javadoc +1. The javadoc tool did not generate any warning messages.

javac +1. The applied patch does not generate any new compiler warnings.

findbugs +1. The patch does not introduce any new Findbugs warnings.

core tests -1. The patch failed core unit tests.

contrib tests -1. The patch failed contrib unit tests.

Test results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1019/testReport/
Findbugs warnings: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1019/artifact/trunk/build/test/findbugs/newPatchFindbugsWarnings.html
Checkstyle results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1019/artifact/trunk/build/test/checkstyle-errors.html
Console output: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1019/console

This message is automatically generated.

[ Permlink | « Hide ]

Owen O'Malley added a comment - 02/Nov/07 06:25 AM

The JavaDoc for TupleWritable's class description isn't right. (It wasn't updated.)

The IOExceptions that are wrapping other errors should have descriptive string messages in them.

You don't need to define a new ReflectionUtils.newInstance without the config, because if you pass in null, it won't use it.

All of the instances the use cls.newInstance should be using ReflectionUtils.newInstance, since it does the constructor cache and handles the non-public class/constructor problems.

The fieldnames m and n need descriptive names.

I really don't like protected fields, especially when they are set/used multiple levels below where they are defined.

[ Permlink | « Hide ]

Joydeep Sen Sarma added a comment - 05/Nov/07 11:13 PM

Chris - can you help me understand how the splits work? This might be useful for some of our apps and trying to understand what assumptions etc. are being made.

we have sorted data files containing same sets of keys - but corresponding hdfs chunks of each file may not have the same set of keys. it wasnt clear to me from going through the patch how the merge-join is being parallelized. from node.getsplits - it seemed as if the ith split of the join record reader is composed of the ith split of each of the component files. but in this case - the join keys wouldn't line up ..

also - given that the map task works on multiple hdfs files - where does it get scheduled?

[ Permlink | « Hide ]

Chris Douglas added a comment - 05/Nov/07 11:55 PM

Joydeep-

The assumption it makes is precisely as you describe it: the ith split from each source must contain the same keys. It does only the most rudimentary verification of this, IIRC verifying that it received an equal number of splits from each source. Generally, getting splits should be cheap, so it doesn't verify key ranges for any of the splits (and probably ought not to).

I've asked around, and "the way" out of this onerous constraint involves using MapFiles. At a high level, you need an index for your input data so your splits can be informed. I'm not familiar with the details here, I'm afraid.

CompositeInputSplit::getLocations() returns an unweighted union of hosts from its child splits. It would be preferable to weight a host that contains multiple splits for a given composite split, but for now it provides a flat list.

[ Permlink | « Hide ]

Joydeep Sen Sarma added a comment - 06/Nov/07 12:50 AM

understood. i was thinking this might be using mapfiles or some kind of binary search to line up splits.

Dumb question - our data is laid out as files (representing partitions) within a single directory - with a directory representing a pseudo-table. Is this compatible with where u are going? ie. - can i represent join one such directory against another - with (say) an inputformat that emits each file as a split (and making sure the order is the same)?

The other case is that sometimes one dataset is partitioned (say) 16 way - but another is partitioned 32 way. This can happen when datasets are of unequal size (otherwise we end up creating too many files). In the above case, 2 files from the latter dataset have to be joined against each file from the former (assuming simple modulo arithmetic partitioning). would this be possible?

[ Permlink | « Hide ]

Chris Douglas added a comment - 06/Nov/07 01:30 AM

I don't know if this is helpful, but: as it exists now, the framework is incapable of finer granularity than an InputFormat, but neither will it object whatever you can fit into that framework.

What you describe- directories as pseudo-tables with files as partitions- sounds like exactly what this is geared toward.

As an example of a workaround/partial fit, consider your 16/32 way case. Whether it would be worthwhile/possible to express in the existing code will depend on a few factors: if the two files you're joining in the 32-way set are pairwise disjoint, then you can simply use an OverrideRecordReader with two custom InputFormats (each taking one "half" of the pair) to "join" them. However, if they're not disjoint, then you'll lose values. ¹ Feeding the output of that into a join with your 16 way dataset might work, but it's a bit of a hack. You'd need to be certain of the partitions of both datasets to be confident in your results.

Notes
1. Really, you're looking for a different implementation of CompositeRecordReader::JoinCollector that emits values from each source in turn, rather than emitting the cross-product; this is being considered, but may not be in the immediate future. It's of limited use with the requirement that each source be sorted an partitioned in the same way, unfortunately. Most simply want to merge two sorted datasets without worrying about how they're partitioned (HADOOP-2120).

[ Permlink | « Hide ]

Chris Douglas added a comment - 09/Nov/07 01:34 AM

Updated javadocs, made better use of ReflectionUtils, improved some variable names, made protected fields private or final (excluding the parser, which is temporary).

[ Permlink | « Hide ]

Hadoop QA added a comment - 09/Nov/07 03:03 AM

+1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12369206/2085-4.patch
against trunk revision r592860.

@author +1. The patch does not contain any @author tags.

javadoc +1. The javadoc tool did not generate any warning messages.

javac +1. The applied patch does not generate any new compiler warnings.

findbugs +1. The patch does not introduce any new Findbugs warnings.

core tests +1. The patch passed core unit tests.

contrib tests +1. The patch passed contrib unit tests.

Test results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1077/testReport/
Findbugs warnings: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1077/artifact/trunk/build/test/findbugs/newPatchFindbugsWarnings.html
Checkstyle results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1077/artifact/trunk/build/test/checkstyle-errors.html
Console output: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1077/console

This message is automatically generated.

[ Permlink | « Hide ]

Milind Bhandarkar added a comment - 15/Nov/07 08:33 PM

A few comments on the patch:

If mapred.join.expr is not specified, CompositeInputFormat should throw a better exception,, rather than NPE.

A simple benchmark that we can use to compare performance with with the reducer-side joins is desirable. But, it can be a separate jira.

The motivation and design that is in this jira should be in package.html for o.a.h.mapred.join.

[ Permlink | « Hide ]

Chris Douglas added a comment - 16/Nov/07 09:42 PM

Addressed NPE and included interface info from this JIRA in o/a/h/mapred/join/package.html

[ Permlink | « Hide ]

Hadoop QA added a comment - 17/Nov/07 08:16 AM

-1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12369694/2085-5.patch
against trunk revision r595563.

@author +1. The patch does not contain any @author tags.

javadoc +1. The javadoc tool did not generate any warning messages.

javac +1. The applied patch does not generate any new compiler warnings.

findbugs +1. The patch does not introduce any new Findbugs warnings.

core tests +1. The patch passed core unit tests.

contrib tests -1. The patch failed contrib unit tests.

Test results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1110/testReport/
Findbugs warnings: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1110/artifact/trunk/build/test/findbugs/newPatchFindbugsWarnings.html
Checkstyle results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1110/artifact/trunk/build/test/checkstyle-errors.html
Console output: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1110/console

This message is automatically generated.

[ Permlink | « Hide ]

Chris Douglas added a comment - 17/Nov/07 09:34 PM

Trying hudson again

[ Permlink | « Hide ]

Hadoop QA added a comment - 17/Nov/07 11:03 PM

+1 overall. Here are the results of testing the latest attachment
http://issues.apache.org/jira/secure/attachment/12369694/2085-5.patch
against trunk revision r595563.

@author +1. The patch does not contain any @author tags.

javadoc +1. The javadoc tool did not generate any warning messages.

javac +1. The applied patch does not generate any new compiler warnings.

findbugs +1. The patch does not introduce any new Findbugs warnings.

core tests +1. The patch passed core unit tests.

contrib tests +1. The patch passed contrib unit tests.

Test results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1116/testReport/
Findbugs warnings: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1116/artifact/trunk/build/test/findbugs/newPatchFindbugsWarnings.html
Checkstyle results: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1116/artifact/trunk/build/test/checkstyle-errors.html
Console output: http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Patch/1116/console

This message is automatically generated.

[ Permlink | « Hide ]

Owen O'Malley added a comment - 07/Dec/07 10:06 PM

I just committed this. Thanks, Chris.

[ Permlink | « Hide ]

Hudson added a comment - 08/Dec/07 12:17 PM

Integrated in Hadoop-Nightly #326 (See http://lucene.zones.apache.org:8080/hudson/job/Hadoop-Nightly/326/)

[ Permlink | « Hide ]

Joydeep Sen Sarma added a comment - 04/Mar/08 06:01 PM

Hey folks - i was thinking of a easier way to do merge joins that removes the need for any assumptions about equipartitioning. it can also work with sorted and non-sorted data sets. It's also extremely simple to implement:

assume that there are two (or more) tables to be joined. Let's say Table A is sorted on join column and Table B is not.
we use Hadoop-2921 (or same technique implemented at higher layer) to force maps on A to be aligned by sort boundary. We also turn off sorting in map phase when maps work on table A. B is treated as usual.
Reducers do a merge join of A and B.

If A and B are both sorted by the join column - then we are doing a pure merge-join in the reducer (maps will not sort).

A and B don't need to be equipartitioned. It's of course not as efficient as the merge-join implemented here - but its also way more flexible .. (metadata about table sorting columns could be maintained outside and the map jobs configured based on whether sort and join columns match).

[ Permlink | « Hide ]

Chris Douglas added a comment - 04/Mar/08 10:04 PM

Joydeep- excluding the optimization for not re-sorting A, it sounds like you're describing the join framework in contrib. The idea of metadata storing sorting columns, etc. is compelling, but a reasonable use of it would best be done by something like Pig, no? The most likely next step for more complex, reduce-side joins would be different map tasks for different datasets (e.g. emit result of operation w/ col 1,3 in B; identity for A, possibly in different formats sorted on whatever) followed by a join in the reduce. A sufficiently general execution engine- that could made decisions about whether or not the data is already sorted on some column, whether the join can happen on the map or reduce side, etc- belongs in framework code, I agree, but I'm less convinced it should live in this framework code.

We could change the idea of a job to include map tasks across multiple datasets- similar, yet very much unlike the join work in this patch- followed by a reduce step. To take your example, starting two different maps over A and B st the partitions are congruent (i.e. K1 in A and K1 in B go to the same partition), essentially providing different map classes for different input paths. Of course, all good ideas have JIRAs: ~~HADOOP-372~~