My spark application process the files (average size is 20 MB) with custom hadoop input format and stores the result in HDFS.
Following is the code snippet.
Configuration conf = new Configuration();
JavaPairRDD<Text, Text> baseRDD = ctx
.newAPIHadoopFile(input, CustomInputFormat.class,Text.class, Text.class, conf);
JavaRDD<myClass> mapPartitionsRDD = baseRDD
.mapPartitions(new FlatMapFunction<Iterator<Tuple2<Text, Text>>, myClass>() {
//my logic goes here
}
//few more translformations
result.saveAsTextFile(path);
This application creates 1 task/ partition per file and processes and stores the corresponding part file in HDFS.
i.e, For 10,000 input files 10,000 tasks are created and 10,000 part files are stored in HDFS.
Both mapPartitions and map operations on baseRDD are creating 1 task per file.
SO question
How to set the number of partitions for newAPIHadoopFile?
suggests to set
conf.setInt("mapred.max.split.size", 4); for configuring no of partitions.
But when this parameter is set CPU is utilized at maximum and none of the stage is not started even after long time.
If I don't set this parameter then application will be completed successfully as mentioned above.
How to set number of partitions with newAPIHadoopFile and increase the efficiency?
What happens with mapred.max.split.size option?
============
update:
What happens with mapred.max.split.size option?
In my use case file size is small and changing the split size options are irrelevant here.
more info on this SO: Behavior of the parameter "mapred.min.split.size" in HDFS
Just use baseRDD.repartition(<a sane amount>).mapPartitions(...). That will move the resulting operation to fewer partitions, especially if your files are small.
Related
I am trying to run a script or a custom processor to group data by given attributes every hour. Queue size is up to 30-40k on a single run and it might go up to 200k depending on the case.
MergeContent does not fit since there is no limit on min-max counts.
RouteOnAttribute does not fit since there are too many combinations.
Solution 1: Consume all flow files and group by attributes and create the new flow file and push the new one. Not ideal but gave it a try.
While running this when I had 33k flow files on queue waiting.
session.getQueueSize().getObjectCount()
This number is returning 10k all the time even though I increased the queue threshold numbers on output flows.
Solution 2: Better approach is consume one flow file and and filter flow files matching the provided attributes
final List<FlowFile> flowFiles = session.get(file -> {
if (correlationId.equals(Arrays.stream(keys).map(file::getAttribute).collect(Collectors.joining(":"))))
return FlowFileFilter.FlowFileFilterResult.ACCEPT_AND_CONTINUE;
return FlowFileFilter.FlowFileFilterResult.REJECT_AND_CONTINUE;
});
Again with 33k waiting in the queue I was expecting around 200 new grouped flow files but 320 is created. It looks like a similar issue above and does not scan all waiting flow files on filter query.
Problems-Question:
Is there a parameter to change so this getObjectCount can take up to 300k?
Is there a way to filter all waiting flow files again by changing a parameter or by changing the processor?
I tried making default queue threshold 300k on nifi.properties but it didn't help
in nifi.properties there is a parameter that affects batching behavior
nifi.queue.swap.threshold=20000
here is my test flow:
1. GenerateFlowFile with "batch size = 50K"
2. ExecuteGroovyScript with script below
3. LogAttrribute (disabled) - just to have queue after groovy
groovy script:
def ffList = session.get(100000) // get batch with maximum 100K files from incoming queue
if(!ffList)return
def ff = session.create() // create new empty file
ff.batch_size = ffList.size() // set attribute to real batch size
session.remove(ffList) // drop all incoming batch files
REL_SUCCESS << ff // transfer new file to success
with parameters above there are 4 files generated in output:
1. batch_size = 20000
2. batch_size = 10000
3. batch_size = 10000
4. batch_size = 10000
according to documentation:
There is also the notion of "swapping" FlowFiles. This occurs when the number of FlowFiles in a connection queue exceeds the value set in the nifi.queue.swap.threshold property. The FlowFiles with the lowest priority in the connection queue are serialized and written to disk in a "swap file" in batches of 10,000.
This explains that from 50K incoming files - 20K it keeps inmemory and others in swap batched by 10K.
i don't know how increasing of nifi.queue.swap.threshold property will affect your system performance and memory consumption, but i set it to 100K on my local nifi 1.16.3 and it looks good with multiple small files, and first batch increased to 100K by this.
I'm investigating the performances of a Flink job that transports data from Kafka to an S3 Sink.
We are using a BucketingSink to write parquet files. The bucketing logic divides the messages having a folder per type of data, tenant (customer), date-time, extraction Id, etc etc. This results in each file is stored in a folder structure composed by 9-10 layers (s3_bucket:/1/2/3/4/5/6/7/8/9/myFile...)
If the data is distributed as bursts of messages for tenant-type we see good performances in writing, but when the data is more a white noise distribution on thousands of tenants, dozens of data types and multiple extraction IDs, we have an incredible loss of performances. (in the order of 300x times)
Attaching a debugger, it seems the issue is connected to the number of handlers open at the same time on S3 to write data. More specifically:
Researching in the hadoop libraries used to write to S3 I have found some possible improvements setting:
<name>fs.s3a.connection.maximum</name>
<name>fs.s3a.threads.max</name>
<name>fs.s3a.threads.core</name>
<name>fs.s3a.max.total.tasks</name>
But none of these made a big difference in throughput.
I also tried to flatten the folder structure to write to a single key like (1_2_3_...) but also this didn't bring any improvement.
Note: The tests have been done on Flink 1.8 with the Hadoop FileSystem (BucketingSink), writing to S3 using the hadoop fs libraries 2.6.x (as we use Cloudera CDH 5.x for savepoints), so we can't switch to StreamingFileSink.
After the suggestion from Kostas in https://lists.apache.org/thread.html/50ef4d26a1af408df8d9abb70589699cb6b26b2600ab6f4464e86ea4%40%3Cdev.flink.apache.org%3E
The culprit of the slow-down is this piece of code:
https://github.com/apache/flink/blob/master/flink-connectors/flink-connector-filesystem/src/main/java/org/apache/flink/streaming/connectors/fs/bucketing/BucketingSink.java#L543-L551
This alone takes around 4-5 secs, with a total of 6 secs to open the file. Logs from an instrumented call:
2020-02-07 08:51:05,825 INFO BucketingSink - openNewPartFile FS verification
2020-02-07 08:51:09,906 INFO BucketingSink - openNewPartFile FS verification - done
2020-02-07 08:51:11,181 INFO BucketingSink - openNewPartFile FS - completed partPath = s3a://....
This together with the default setup of the bucketing sink with 60 secs inactivity rollover
https://github.com/apache/flink/blob/master/flink-connectors/flink-connector-filesystem/src/main/java/org/apache/flink/streaming/connectors/fs/bucketing/BucketingSink.java#L195
means that with more than 10 parallel bucket on a slot by the time we finish creating the last bucket the first one became stale, so needs to be rotated generating a blocking situation.
We solved this by replacing the BucketingSink.java and deleting the FS check mentioned above:
LOG.debug("Opening new part file FS verification");
if (!fs.exists(bucketPath)) {
try {
if (fs.mkdirs(bucketPath)) {
LOG.debug("Created new bucket directory: {}", bucketPath);
}
}
catch (IOException e) {
throw new RuntimeException("Could not create new bucket path.", e);
}
}
LOG.debug("Opening new part file FS verification - done");
as we see that the sink works fine without it, now the file opening takes ~1.2sec.
Moreover we set the default inactive threshold to 5 mins. With this changes we can easily handle more than 200 buckets per slot (once the job takes speed it will ingest on all the slots so postponing the inactive timeout)
I am using Flume to store sensor data in HDFS. Once the data is received through MQTT. The subscriber posts the data in JSON format to Flume HTTP listener. It is currently working fine, but the problem is that flume is not writing to HDFS file till I stop it (or the size of the file reachs 128MB). I am using Hive to apply a schema on read. Unfortunately, the resulting hive table contains only 1 entry. This is normal because Flume did not write new coming data to file (loaded by Hive).
Is there any manner to force Flume to write new coming data to HDFS in a near-real time way? So, I don't need to restart it or to use small files?
here is my flume configuration:
# Name the components on this agent
emsFlumeAgent.sources = http_emsFlumeAgent
emsFlumeAgent.sinks = hdfs_sink
emsFlumeAgent.channels = channel_hdfs
# Describe/configure the source
emsFlumeAgent.sources.http_emsFlumeAgent.type = http
emsFlumeAgent.sources.http_emsFlumeAgent.bind = localhost
emsFlumeAgent.sources.http_emsFlumeAgent.port = 41414
# Describe the sink
emsFlumeAgent.sinks.hdfs_sink.type = hdfs
emsFlumeAgent.sinks.hdfs_sink.hdfs.path = hdfs://localhost:9000/EMS/%{sensor}
emsFlumeAgent.sinks.hdfs_sink.hdfs.rollInterval = 0
emsFlumeAgent.sinks.hdfs_sink.hdfs.rollSize = 134217728
emsFlumeAgent.sinks.hdfs_sink.hdfs.rollCount=0
#emsFlumeAgent.sinks.hdfs_sink.hdfs.idleTimeout=20
# Use a channel which buffers events in memory
emsFlumeAgent.channels.channel_hdfs.type = memory
emsFlumeAgent.channels.channel_hdfs.capacity = 10000
emsFlumeAgent.channels.channel_hdfs.transactionCapacity = 100
# Bind the source and sinks to the channel
emsFlumeAgent.sources.http_emsFlumeAgent.channels = channel_hdfs
emsFlumeAgent.sinks.hdfs_sink.channel = channel_hdfs
I think the tricky bit here is that you would like to write data to HDFS in near real time but don't want small files either (for obvious reasons) and this could be a difficult thing to a achieve.
You'll need to find optimal balance between the following two parameters:
hdfs.rollSize (Default = 1024) - File size to trigger roll, in bytes (0: never roll based on file size)
and
hdfs.batchSize (Default = 100) - Number of events written to file before it is flushed to HDFS
If your data is not likely to reach 128 MB in the preferred time duration, then you may need to reduce the rollSize but only to an extent that you don't run into the small files problem.
Since, you have not set any batch size in your HDFS sink, you should see the results of HDFS flush after every 100 records but once the size of the flushed records jointly reaches 128 MB, the contents would be rolled up in a 128 MB file. Is this also not happening? Could you please confirm?
Hope this helps!
I was trying loading file using hadoop API as an experiment.
I want to set replication to minimum as this one is for experiment.
I first tried this with FileSystem.setReplication():
Configuration config = new Configuration();
config.set("fs.defaultFS","hdfs://192.168.248.166:8020");
FileSystem dfs2 = FileSystem.get(config);
Path src2 = new Path("C:\\Users\\abc\\Desktop\\testfile.txt");
Path dst2 = new Path(dfs2.getWorkingDirectory()+"/tempdir");
dfs2.copyFromLocalFile(src2, dst2);
dfs2.setReplication(dst2, (short)1); /**setting replication**/
The replica was shown as 1, but it was available on 3 datanodes.
When I tried it with Configuration.set():
Configuration config = new Configuration();
config.set("fs.defaultFS","hdfs://192.168.248.166:8020");
config.set("dfs.replication", "1"); /**setting replication**/
FileSystem dfs2 = FileSystem.get(config);
Path src2 = new Path("C:\\Users\\abc\\Desktop\\testfile.txt");
Path dst2 = new Path(dfs2.getWorkingDirectory()+"/tempdir");
This gave the desired outcome (1 replica available on 1 datanode)
Why there are two APIs for the same thing?
What is the difference between these two?
The difference is that Filesystem's setReplication() sets the replication of an existing file on HDFS. In your case, you first copy the local file testFile.txt to HDFS, using the default replication factor (3) and then change the replication factor of this file to 1. After this command, it takes a while until the over-replicated blocks get deleted. (source)
On the other hand, when you use the config.set("dfs.replication", "1"); command to set the replication, you can copy the local file after that, so its blocks get copied just once, from the first time.
In other words, I believe (but I might be wrong) that both commands have the same final result, but you have to wait a little bit until the first one is carried out.
I am trying to write the flume events in Amaozn S3.The events written in S3 is in compressed format. My Flume configuration is given below. I am facing a data loss. Based on the configuration given below, if I publish 20000 events, I receive only 1000 events and all other data is lost. But When I disable the rollcount, rollSize and rollInterval configurations, all the events are received but there are 2000 small files created. Is there any wrong in my configuration settings? Should I add any other configurations?
injector.sinks.s3_3store.type = hdfs
injector.sinks.s3_3store.channel = disk_backed4
injector.sinks.s3_3store.hdfs.fileType = CompressedStream
injector.sinks.s3_3store.hdfs.codeC = gzip
injector.sinks.s3_3store.hdfs.serializer = TEXT
injector.sinks.s3_3store.hdfs.path = s3n://CID:SecretKey#bucketName/dth=%Y-%m-%d-%H
injector.sinks.s3_1store.hdfs.filePrefix = events-%{receiver}
# Roll when files reach 256M or after 10m, whichever comes first
injector.sinks.s3_3store.hdfs.rollCount = 0
injector.sinks.s3_3store.hdfs.idleTimeout = 600
injector.sinks.s3_3store.hdfs.rollSize = 268435456
#injector.sinks.s3_3store.hdfs.rollInterval = 3600
# Flush data to buckets every 1k events
injector.sinks.s3_3store.hdfs.batchSize = 10000
For starters: if you disable your setting for rollCount, rollSize and so on, flume will revert to defaults, hence the small files you receive, those are the default.
The relevant aspect is this:
injector.sinks.s3_3store.hdfs.batchSize = 10000
it basically tells your sink to collect 10.000 events before flushing. If you reduce that amount, you'll get smaller files too, because S3 in contrast to regular HDFS doesn't support file appends. Once you flush, the files will be closed and a new file will be created.
Try to determine which amount of events your sink will receive within a short time frame of a couple of minutes or so and set that value as you batch size.