Details
-
New Feature
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
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).