hadoop-pig-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Gianmarco De Francisci Morales (JIRA)" <j...@apache.org>
Subject [jira] Updated: (PIG-1295) Binary comparator for secondary sort
Date Sat, 17 Jul 2010 10:18:50 GMT

     [ https://issues.apache.org/jira/browse/PIG-1295?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Gianmarco De Francisci Morales updated PIG-1295:
------------------------------------------------

    Attachment: PIG-1295_0.9.patch

Implemented a new compareBinInterSedesTuple() method.
This new method works with data serialized with [PIG-1472] format.
It relies on DataType for data type comparison by translating the data types.
I implemented some logic to read unsinged ints from bytes and shorts because I am using ByteBuffer
that does not implement the DataInput interface. We might want to change that later.
For now complex data types cause the whole tuple to be deserialized, but I have put some placeholders
for methods to deserialize single complex objects and continue the normal execution flow.
I plan to fill them in when I move the code inside BinInterSedes so that I can use the methods
to read complex data types (readBag, readMap, etc..). This will involve some juggling around
between DataInputBuffer and ByteBuffer (this is why we might consider to switch out ByteBuffer)
to get the cursors consistent among them.

I think the next step would be splitting the comparator and putting it into the serialization
class (BinInterSedes). I don't know yet how to modify the class interface to make the comparator
class available outside (but I think this relates to phase 2).

I see no good place to put the class that implements the comparator for DefaultTuple, and
Daniel said it is a minor case, so should I just throw away the code for DefaultTuple?

> Binary comparator for secondary sort
> ------------------------------------
>
>                 Key: PIG-1295
>                 URL: https://issues.apache.org/jira/browse/PIG-1295
>             Project: Pig
>          Issue Type: Improvement
>          Components: impl
>    Affects Versions: 0.7.0
>            Reporter: Daniel Dai
>            Assignee: Gianmarco De Francisci Morales
>             Fix For: 0.8.0
>
>         Attachments: PIG-1295_0.1.patch, PIG-1295_0.2.patch, PIG-1295_0.3.patch, PIG-1295_0.4.patch,
PIG-1295_0.5.patch, PIG-1295_0.6.patch, PIG-1295_0.7.patch, PIG-1295_0.8.patch, PIG-1295_0.9.patch
>
>
> When hadoop framework doing the sorting, it will try to use binary version of comparator
if available. The benefit of binary comparator is we do not need to instantiate the object
before we compare. We see a ~30% speedup after we switch to binary comparator. Currently,
Pig use binary comparator in following case:
> 1. When semantics of order doesn't matter. For example, in distinct, we need to do a
sort in order to filter out duplicate values; however, we do not care how comparator sort
keys. Groupby also share this character. In this case, we rely on hadoop's default binary
comparator
> 2. Semantics of order matter, but the key is of simple type. In this case, we have implementation
for simple types, such as integer, long, float, chararray, databytearray, string
> However, if the key is a tuple and the sort semantics matters, we do not have a binary
comparator implementation. This especially matters when we switch to use secondary sort. In
secondary sort, we convert the inner sort of nested foreach into the secondary key and rely
on hadoop to sorting on both main key and secondary key. The sorting key will become a two
items tuple. Since the secondary key the sorting key of the nested foreach, so the sorting
semantics matters. It turns out we do not have binary comparator once we use secondary sort,
and we see a significant slow down.
> Binary comparator for tuple should be doable once we understand the binary structure
of the serialized tuple. We can focus on most common use cases first, which is "group by"
followed by a nested sort. In this case, we will use secondary sort. Semantics of the first
key does not matter but semantics of secondary key matters. We need to identify the boundary
of main key and secondary key in the binary tuple buffer without instantiate tuple itself.
Then if the first key equals, we use a binary comparator to compare secondary key. Secondary
key can also be a complex data type, but for the first step, we focus on simple secondary
key, which is the most common use case.
> We mark this issue to be a candidate project for "Google summer of code 2010" program.


-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message