flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chesnay Schepler <ches...@apache.org>
Subject Re: Collector.collect
Date Tue, 02 May 2017 13:39:39 GMT
In the Batch API only a single operator can be chained to another operator.

So we're starting with this code:

    input = ...
    input.filter(conditionA).output(formatA)
    input.filter(conditonB).output(formatB)

In the Batch API this would create a CHAIN(filterA -> formatA) and a 
CHAIN(filterB -> formatB), both having "input" as their input.
Since the filtering is not done as part of "input" the entire input 
DataSet must be sent to both tasks.
This means that both chains have to deserialize the entire DataSet to 
apply the filter; the serialization should only be done once though.

In contrast the solution you wrote creates a single CHAIN(input, 
format), with no serialization in between at all.

The Streaming API doesn't have this limitation and would get by without 
any serialization as well. Probably.

On 02.05.2017 15:23, Newport, Billy wrote:
>
> Why doesn’t this work with batch though. We did
>
> input = ...
> input.filter(conditionA).output(formatA)
> input.filter(conditonB).output(formatB)
>
> And it was pretty slow compared with a custom outputformat with an 
> integrated filter.
>
> *From:*Chesnay Schepler [mailto:chesnay@apache.org]
> *Sent:* Monday, May 01, 2017 12:56 PM
> *To:* Newport, Billy [Tech]; 'user@flink.apache.org'
> *Subject:* Re: Collector.collect
>
> Oh you have multiple different output formats, missed that.
>
> For the Batch API you are i believe correct, using a custom 
> output-format is the best solution.
>
> In the Streaming API the code below should be equally fast, if the 
> filtered sets don't overlap.
>
> input = ...
> input.filter(conditionA).output(formatA)
> input.filter(conditonB).output(formatB)
>
> That is because all filters would be chained; hell all sources might 
> be as well (not to sure on this one).
>
> On 01.05.2017 17:05, Newport, Billy wrote:
>
>     There is likely a bug then, the ENUM,Record stream to a filter to
>     a set of outputformats per filter was slower than the
>     BITMASK,Record to single OutputFormat which demux’s the data to
>     each file internally
>
>     Are you saying do a custom writer inside a map rather than either
>     of the 2 above approaches?
>
>     *From:*Chesnay Schepler [mailto:chesnay@apache.org]
>     *Sent:* Monday, May 01, 2017 10:41 AM
>     *To:* user@flink.apache.org <mailto:user@flink.apache.org>
>     *Subject:* Re: Collector.collect
>
>     Hello,
>
>     @Billy, what prevented you from duplicating/splitting the record,
>     based on the bitmask, in a map function before the sink?
>     This shouldn't incur any serialization overhead if the sink is
>     chained to the map. The emitted Tuple could also share the
>     GenericRecord; meaning you don't even have to copy it.
>
>     On 01.05.2017 14:52, Newport, Billy wrote:
>
>         We’ve done that but it’s very expensive from a serialization
>         point of view when writing the same record multiple times,
>         each in a different tuple.
>
>         For example, we started with this:
>
>         .collect(new Tuple<Short, GenericRecord)).
>
>         The record would be written with short = 0 and again with
>         short = 1. This results in the GenericRecord being serialized
>         twice. You also prolly need filters on the output dataset
>         which is expensive also.
>
>         We switched instead to a bitmask. Now, we write the record
>         once and set bits in the short for each file the record needs
>         to be written to. Our next step is to write records to a file
>         based on the short. We wrote a new outputrecordformat which
>         checks the bits in the short and writes the GenericRecord to
>         each file for the corresponding bit. This means no filter to
>         split the records for each file and this is much faster.
>
>         We’re finding a need to do this kind of optimization pretty
>         frequently with flink.
>
>         *From:*Gaurav Khandelwal [mailto:gaurav671989@gmail.com]
>         *Sent:* Saturday, April 29, 2017 4:32 AM
>         *To:* user@flink.apache.org <mailto:user@flink.apache.org>
>         *Subject:* Collector.collect
>
>         Hello
>
>         I am working on RichProcessFunction and I want to emit
>         multiple records at a time. To achieve this, I am currently
>         doing :
>
>         while(condition)
>
>         {
>
>          Collector.collect(new Tuple<>...);
>
>         }
>
>         I was wondering, is this the correct way or there is any other
>         alternative.
>


Mime
View raw message