Hi Scott,
thanks for all the suggestions. I really appreciate your support.
Unfortunately, I could not solve the problem so far.
That's what I have tried:
1. Switched to UTF8 everywhere, including changing the interface to <Utf8,
SomeSpecificJavaClass>
2. Always generate new instances before collecting (new Utf8("fromString") for
the key, clone for the value)
The problem persists - records seem to get lost between mapper and reducer.
Interestingly, it's only reproducible with large datasets. So, if I run a
relatively small set of 6 million input rows, I do not get any differences,
however, on a 10 million input dataset the difference shows up:
Map input records: 10,000,000
Map input bytes: 11,458,340,172
Map output bytes: 30,420,106,592
Map output records: 28,196,842
Reduce input records: 28,053,314
I'm trying to simplify the job further.
Do you have any further ideas?
Thanks,
Vyacheslav
On Aug 17, 2011, at 7:18 PM, Scott Carey wrote:
> On 8/17/11 5:02 AM, "Vyacheslav Zholudev" <[email protected]>
> wrote:
>
>> btw,
>>
>> I was thinking to try it with Utf8 objects instead of strings and I wanted
>> to reuse the same Utf8 object instead of creating new from String upon each
>> map() call.
>> Why does not the Utf8 class have a method for setting bytes via a String
>> object?
>
>
> We could add that, but it won't help performance much in this case since the
> performance improvement from reuse has more to do with the underlying byte[]
> than the Utf8 object.
> The expensive part of String is the conversion from an underlying char[] to a
> byte[] (Utf8.getBytesFor()), so this would not help much. It would probably
> be faster to use String directly rather than wrap it with Utf8 each time.
>
> Rather than have a static method like the below, I would propose that an
> instance method be made that does the same thing, something like
>
> public void setValue(String val) {
> // gets bytes, replaces private byte array, replaces cached string — no
> system array copy.
> }
>
> which would be much more efficient.
>
>
>>
>> I created the following code snippet:
>>
>> public static Utf8 reuseUtf8Object(Utf8 container, String strToReuse) {
>> byte[] strBytes = Utf8.getBytesFor(strToReuse);
>> container.setByteLength(strBytes.length);
>> System.arraycopy(strBytes, 0, container.getBytes(), 0,
>> strBytes.length);
>> return container;
>> }
>>
>> Would that be useful if this code is encapsulated into the Utf8 class?
>>
>> Best,
>> Vyacheslav
>>
>> On Aug 17, 2011, at 3:56 AM, Scott Carey wrote:
>>
>>> On 8/16/11 3:56 PM, "Vyacheslav Zholudev" <[email protected]>
>>> wrote:
>>>
>>>> Hi, Scott,
>>>>
>>>> thanks for your reply.
>>>>
>>>>> What Avro version is this happening with? What JVM version?
>>>>
>>>> We are using Avro 1.5.1 and Sun JDK 6, but the exact version I will have
>>>> to look up.
>>>>
>>>>>
>>>>> On a hunch, have you tried adding -XX:-UseLoopPredicate to the JVM args
>>>>> if
>>>>> it is Sun and JRE 6u21 or later? (some issues in loop predicates affect
>>>>> Java 6 too, just not as many as the recent news on Java7).
>>>>>
>>>>> Otherwise, it may likely be the same thing as AVRO-782. Any extra
>>>>> information related to that issue would be welcome.
>>>>
>>>> I will have to collect it. In the meanwhile, do you have any reasonable
>>>> explanations of the issue besides it being something like AVRO-782?
>>>
>>> What is your key type (map output schema, first type argument of Pair)?
>>> Is your key a Utf8 or String? I don't have a reasonable explanation at
>>> this point, I haven't looked into it in depth with a good reproducible
>>> case. I have my suspicions with how recycling of the key works since Utf8
>>> is mutable and its backing byte[] can end up shared.
>>>
>>>
>>>
>>>>
>>>> Thanks a lot,
>>>> Vyacheslav
>>>>
>>>>>
>>>>> Thanks!
>>>>>
>>>>> -Scott
>>>>>
>>>>>
>>>>>
>>>>> On 8/16/11 8:39 AM, "Vyacheslav Zholudev"
>>>>> <[email protected]>
>>>>> wrote:
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> I'm having multiple hadoop jobs that use the avro mapred API.
>>>>>> Only in one of the jobs I have a visible mismatch between a number of
>>>>>> map
>>>>>> output records and reducer input records.
>>>>>>
>>>>>> Does anybody encountered such a behavior? Can anybody think of possible
>>>>>> explanations of this phenomenon?
>>>>>>
>>>>>> Any pointers/thoughts are highly appreciated!
>>>>>>
>>>>>> Best,
>>>>>> Vyacheslav
>>>>>
>>>>>
>>>>
>>>> Best,
>>>> Vyacheslav
>>>>
>>>>
>>>>
>>>
>>>
>>
