flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Newport, Billy" <Billy.Newp...@gs.com>
Subject RE: Flink parquet read.write performance
Date Thu, 24 Aug 2017 12:42:48 GMT
If we use two sinks with the same folder then we get file name collisions between the two sinks.

It sounds like even if we did that, flink isn’t capable of chaining it regardless, no?

We find ourselves having to manually optimize the data flow quite a bit to tell you the truth.
For example:

FileDataset -> [Filter True -> Flow;Filter False -> Flow 2]

Is slower than reading the file twice i.e. FileDataset -> Filter True -> Flow and another
flow FileDataset -> Filter False -> Flow 2.

From: Stephan Ewen [mailto:sewen@apache.org]
Sent: Wednesday, August 23, 2017 12:21 PM
To: Aljoscha Krettek
Cc: Newport, Billy [Tech]; Chan, Regina [Tech]; user@flink.apache.org
Subject: Re: Flink parquet read.write performance


The sink is merely a union of the result of the co-group and the one data source.
Can't you just make to distinct pipelines out of that? One with co-group -> data sink pipeline
and one with the source->sink pipeline?
They could even be part of the same job...


On Wed, Aug 23, 2017 at 5:51 PM, Aljoscha Krettek <aljoscha@apache.org<mailto:aljoscha@apache.org>>

The reason is that there are two (or more) different Threads doing the reading. As an illustration,
consider first this case:

DataSet input = ...
input.map(new MapA()).map(new MapB())

Here, MapB is technically "wrapped" by MapA and when MapA emits data this is directly going
the the map() method of MapB. The two functions are chained.

Now, in this other case the methods cannot be chained:

DataSet input1= ...
DataSet input2
DataSet mappedA = input1.map(new MapA())
DataSet mappedB = input2.map(new MapB())

mappedA.union(mappedB).map(new MapC())

Here, there is (at least) one thread per map because none of MapA or MapB could wrap MapC
such that the other one (either MapA or MapB) can still send data into MapC. Data is sent
across a channel between the Threads and whenever that happens the data is serialised.

Technically, we could avoid serialization if we knew that two Threads are running in the same
JVM but this is not something that Flink currently does.


On 23. Aug 2017, at 17:12, Newport, Billy <Billy.Newport@gs.com<mailto:Billy.Newport@gs.com>>

Thanks Aljoscha for the prompt response.

Can you explain the technical reason for the single predecessor rule? This makes what we are
trying to do much more expensive. Really what we’re doing is reading a parquet file, doing
several maps/filters on the records and writing to the parquet. There is no serialization
besides the parquet operations needed at all. The current flink implementation adds an expensive
serialize/deserialize for no apparent purpose in the code.


From: Aljoscha Krettek [mailto:aljoscha@apache.org]
Sent: Saturday, August 19, 2017 1:45 AM
To: Chan, Regina [Tech]
Cc: Newport, Billy [Tech]; user@flink.apache.org<mailto:user@flink.apache.org>
Subject: Re: Flink parquet read.write performance


The Sink cannot be chained to the previous two operations because there are two operations.
Chaining only works if there is one predecessor operation. Data transfer should still be pipelined
but you will see serialisation overhead. What kind of TypeSerializer is used at that boundary?

On 18. Aug 2017, at 21:15, Chan, Regina <Regina.Chan@gs.com<mailto:Regina.Chan@gs.com>>

We profiled it and it looks like its sending the output of the datastoure->filter->map->map
to the an intermediate result partition instead of writing directly to the data sink. Because
of this we think it’s slow because it’s spending its time serializing it for no reason.
Why does it do the forward rather than chain to the datasink?



From: Aljoscha Krettek [mailto:aljoscha@apache.org]
Sent: Friday, August 18, 2017 12:14 PM
To: Newport, Billy [Tech]
Cc: user@flink.apache.org<mailto:user@flink.apache.org>
Subject: Re: Flink parquet read.write performance

Hi Billy,

Do you also have the data (picture) from the "Timeline" tab of the completed job? This would
give some hints about how long that other DataSource (with chain) was active. It might be
that the sink is waiting for the other input to become online.


On 18. Aug 2017, at 14:45, Newport, Billy <Billy.Newport@gs.com<mailto:Billy.Newport@gs.com>>


I’m trying to figure out why reading and writing ~5GB worth of parquet files seems to take
3-4 minutes with 10 TaskManagers, 2 slots, 20GB memory, 20 Parallelism. I’ve copied in the
execution plan the taskmanager times below. Other details include that we’re reading 20
snappy compresed parquet files each ~240MB each. (see below)

I’m trying to use this for a milestoning logic where we take new avro files from staging
and join with the existing milestoned parquet data. I have a small staging file with only
about 1500 records inside so I reduce the number of records sent to the cogroup in order to
make this faster. To do this, I’m basically reading in GenericRecords from parquet files
twice, once to filter out for “live” records where we then further filter the records
for ones with keys matching what we found in a separate avro file. This is so reduction of
records makes that part of the plan total to 1 minute 58 secs.

The concern is the other records with non-live/not-matching-keys. In theory, I expect this
to be fast since it’s just chaining the operations across all the way through to the sink.
However, this part takes about 4 minutes. We’re not doing anything different from the other
Datasource aside from mapping a DataSet<GenericRecord> to a Tuple2<Short,GenericRecord>
where the short is a bitmap value mapping to where the record needs to be written.

Other notes:
I checked the backpressure on the datasource->filter->map->map and it was OK. I’m
not sure what else could be holding it up.
I also profiled it when I ran it on a single task manager single slot and it seems to spend
most of the time waiting.

Any ideas? Instead of truly chaining is it writing to disk and serializing multiple times
inside each operation?

Data Source :
hdfs dfs -du -h <folder_name>
240.2 M  <folder_name>/0_partMapper-m-00013.snappy.parquet
237.2 M  <folder_name>/10_partMapper-m-00019.snappy.parquet
241.9 M  <folder_name>/11_partMapper-m-00002.snappy.parquet
243.3 M  <folder_name>/12_partMapper-m-00000.snappy.parquet
238.2 M  <folder_name>/13_partMapper-m-00016.snappy.parquet
241.7 M  <folder_name>/14_partMapper-m-00003.snappy.parquet
241.0 M  <folder_name>/15_partMapper-m-00006.snappy.parquet
240.3 M  <folder_name>/16_partMapper-m-00012.snappy.parquet
240.3 M  <folder_name>/17_partMapper-m-00011.snappy.parquet
239.5 M  <folder_name>/18_partMapper-m-00014.snappy.parquet
237.6 M  <folder_name>/19_partMapper-m-00018.snappy.parquet
240.7 M  <folder_name>/1_partMapper-m-00009.snappy.parquet
240.7 M  <folder_name>/20_partMapper-m-00008.snappy.parquet
236.5 M  <folder_name>/2_partMapper-m-00020.snappy.parquet
242.1 M  <folder_name>/3_partMapper-m-00001.snappy.parquet
241.7 M  <folder_name>/4_partMapper-m-00004.snappy.parquet
240.5 M  <folder_name>/5_partMapper-m-00010.snappy.parquet
241.7 M  <folder_name>/6_partMapper-m-00005.snappy.parquet
239.1 M  <folder_name>/7_partMapper-m-00015.snappy.parquet
237.9 M  <folder_name>/8_partMapper-m-00017.snappy.parquet
240.8 M  <folder_name>/9_partMapper-m-00007.snappy.parquet

yarn-session.sh -nm "delp_uat-IMD_Trading_v1_PROD_PerfTest-REFINER_INGEST"  -jm 4096 -tm 20480
-s 2 -n 10  -d]





View raw message