We have a dataflow streaming job which reads from PubSub, extracts some fields and writes to bigtable. We are observing that dataflow's throughput drops when it is autoscaling. For example, if the dataflow job is currently running with 2 workers and processing at the rate of 100 messages/sec, during autoscaling this rate of 100 messages/sec drops down and some times it drops down to nearly 0 and then increases to 500 messages/sec. We are seeing this every time, dataflow upscales. This is causing higher system lag during autoscaling and bigger spikes of unacknowledged messages in pub/sub.
Is this the expected behavior of dataflow autoscaling or is there a way to maintain this 100 messages/sec while it autoscales and minimize the spices of unacknowledged messages?
(Please Note: 100 messages/sec and 500 messages/sec are just example figures)
job ID: 2017-10-23_12_29_09-11538967430775949506
I am attaching the screen shots of pub/sub stackdriver and dataflow autoscaling.
There is drop in number of pull requests everytime dataflow autoscales. I could not take screenshot with timestamps, but drop in pull requests matches with the time data flow autoscaling.
===========EDIT========================
We are writing to GCS in parallel using below mentioned windowing.
inputCollection.apply("Windowing",
Window.<String>into(FixedWindows.of(ONE_MINUTE))
.triggering(AfterProcessingTime.pastFirstElementInPane()
.plusDelayOf(ONE_MINUTE))
.withAllowedLateness(ONE_HOUR)
.discardingFiredPanes()
)
//Writing to GCS
.apply(TextIO.write()
.withWindowedWrites()
.withNumShards(10)
.to(options.getOutputPath())
.withFilenamePolicy(
new
WindowedFileNames(options.getOutputPath())));
WindowedFileNames.java
public class WindowedFileNames extends FilenamePolicy implements OrangeStreamConstants{
/**
*
*/
private static final long serialVersionUID = 1L;
private static Logger logger = LoggerFactory.getLogger(WindowedFileNames.class);
protected final String outputPath;
public WindowedFileNames(String outputPath) {
this.outputPath = outputPath;
}
#Override
public ResourceId windowedFilename(ResourceId outputDirectory, WindowedContext context, String extension) {
IntervalWindow intervalWindow = (IntervalWindow) context.getWindow();
DateTime date = intervalWindow.maxTimestamp().toDateTime(DateTimeZone.forID("America/New_York"));
String fileName = String.format(FOLDER_EXPR, outputPath, //"orangestreaming",
DAY_FORMAT.print(date), HOUR_MIN_FORMAT.print(date) + HYPHEN + context.getShardNumber());
logger.error(fileName+"::::: File name for the current minute");
return outputDirectory
.getCurrentDirectory()
.resolve(fileName, StandardResolveOptions.RESOLVE_FILE);
}
#Override
public ResourceId unwindowedFilename(ResourceId outputDirectory, Context context, String extension) {
return null;
}
}
What is actually happening is that your throughput is decreasing first, and that is the reason that workers are scaling up.
If you look at your pipeline around 1:30am, the series of events is like so:
Around 1:23am, throughput drops. This builds up the backlog.
Around 1:28am, the pipeline unblocks and starts making progress.
Due to the large backlog, the pipeline scales up to 30 workers.
Also, if you look at the autoscaling UI, the justification for going up to 30 workers is:
"Raised the number of workers to 30 so that the pipeline can catch up
with its backlog and keep up with its input rate."
Hope that helps!
Related
I have an Apache Beam pipeline running on Google Dataflow whose job is rather simple:
It reads individual JSON objects from Pub/Sub
Parses them
And sends them via HTTP to some API
This API requires me to send the items in batches of 75. So I built a DoFn that accumulates events in a list and publish them via this API once they I get 75. This results to be too slow, so I thought instead of executing those HTTP requests in different threads using a thread pool.
The implementation of what I have right now looks like this:
private class WriteFn : DoFn<TheEvent, Void>() {
#Transient var api: TheApi
#Transient var currentBatch: MutableList<TheEvent>
#Transient var executor: ExecutorService
#Setup
fun setup() {
api = buildApi()
executor = Executors.newCachedThreadPool()
}
#StartBundle
fun startBundle() {
currentBatch = mutableListOf()
}
#ProcessElement
fun processElement(processContext: ProcessContext) {
val record = processContext.element()
currentBatch.add(record)
if (currentBatch.size >= 75) {
flush()
}
}
private fun flush() {
val payloadTrack = currentBatch.toList()
executor.submit {
api.sendToApi(payloadTrack)
}
currentBatch.clear()
}
#FinishBundle
fun finishBundle() {
if (currentBatch.isNotEmpty()) {
flush()
}
}
#Teardown
fun teardown() {
executor.shutdown()
executor.awaitTermination(30, TimeUnit.SECONDS)
}
}
This seems to work "fine" in the sense that data is making it to the API. But I don't know if this is the right approach and I have the sense that this is very slow.
The reason I think it's slow is that when load testing (by sending a few million events to Pub/Sub), it takes it up to 8 times more time for the pipeline to forward those messages to the API (which has response times of under 8ms) than for my laptop to feed them into Pub/Sub.
Is there any problem with my implementation? Is this the way I should be doing this?
Also... am I required to wait for all the requests to finish in my #FinishBundle method (i.e. by getting the futures returned by the executor and waiting on them)?
You have two interrelated questions here:
Are you doing this right / do you need to change anything?
Do you need to wait in #FinishBundle?
The second answer: yes. But actually you need to flush more thoroughly, as will become clear.
Once your #FinishBundle method succeeds, a Beam runner will assume the bundle has completed successfully. But your #FinishBundle only sends the requests - it does not ensure they have succeeded. So you could lose data that way if the requests subsequently fail. Your #FinishBundle method should actually be blocking and waiting for confirmation of success from the TheApi. Incidentally, all of the above should be idempotent, since after finishing the bundle, an earthquake could strike and cause a retry ;-)
So to answer the first question: should you change anything? Just the above. The practice of batching requests this way can work as long as you are sure the results are committed before the bundle is committed.
You may find that doing so will cause your pipeline to slow down, because #FinishBundle happens more frequently than #Setup. To batch up requests across bundles you need to use the lower-level features of state and timers. I wrote up a contrived version of your use case at https://beam.apache.org/blog/2017/08/28/timely-processing.html. I would be quite interested in how this works for you.
It may simply be that the extremely low latency you are expecting, in the low millisecond range, is not available when there is a durable shuffle in your pipeline.
I'm trying to understand Beam/Dataflow concepts better, so pretend I have the following streaming pipeline:
pipeline
.apply(PubsubIO.readStrings().fromSubscription("some-subscription"))
.apply(ParDo.of(new DoFn<String, String>() {
#ProcessElement
public void processElement(ProcessContext c) {
String message = c.element();
LOGGER.debug("Got message: {}", message);
c.output(message);
}
}));
How often will the unbounded source pull messages from the subscription? Is this configurable at all (potentially based on windows/triggers)?
Since no custom windowing/triggers have been defined, and there are no sinks (just a ParDo that logs + re-outputs the message), will my ParDo still be executed immediately as messages are received, and is that setup problematic in any way (not having any windows/triggers/sinks defined)?
It will pull messages from the subscription continuously - as soon as a message arrives, it will be processed immediately (modulo network and RPC latency).
Windowing and triggers do not affect this at all - they only affect how the data gets grouped at grouping operations (GroupByKey and Combine). If your pipeline doesn't have grouping operations, windowing and triggers are basically a no-op.
The Beam model does not have the concept of a sink - writing to various storage systems (e.g. writing files, writing to BigQuery etc) is implemented as regular Beam composite transforms, made of ParDo and GroupByKey like anything else. E.g. writing each element to its own file could be implemented by a ParDo whose #ProcessElement opens the file, writes the element to it and closes the file.
I'm trying with Apache Beam 2.1.0 to consume simple data (key,value) from google PubSub and group by key to be able to treat batches of data.
With default trigger my code after "GroupByKey" never fires (I waited 30min).
If I defined custom trigger, code is executed but I would like to understand why default trigger is never fired. I tried to define my own timestamp with "withTimestampLabel" but same issue. I tried to change duration of windows but same issue too (1second, 10seconds, 30seconds etc).
I used command line for this test to insert data
gcloud beta pubsub topics publish test A,1
gcloud beta pubsub topics publish test A,2
gcloud beta pubsub topics publish test B,1
gcloud beta pubsub topics publish test B,2
From documentation it says that we can do one or the other but not necessarily both
If you are using unbounded PCollections, you must use either
non-global windowing OR an aggregation trigger in order to perform a
GroupByKey or CoGroupByKey
It looks to be similar to
Consuming unbounded data in windows with default trigger
Scio: groupByKey doesn't work when using Pub/Sub as collection source
My code
static class Compute extends DoFn<KV<String, Iterable<Integer>>, Void> {
#ProcessElement
public void processElement(ProcessContext c) {
// Code never fires
System.out.println("KEY:" + c.element().getKey());
System.out.println("NB:" + c.element().getValue().spliterator().getExactSizeIfKnown());
}
}
public static void main(String[] args) {
Pipeline p = Pipeline.create(PipelineOptionsFactory.create());
p.apply(PubsubIO.readStrings().fromSubscription("projects/" + args[0] + "/subscriptions/test"))
.apply(Window.into(FixedWindows.of(Duration.standardMinutes(1))))
.apply(
MapElements
.into(TypeDescriptors.kvs(TypeDescriptors.strings(), TypeDescriptors.integers()))
.via((String row) -> {
String[] parts = row.split(",");
System.out.println(Arrays.toString(parts)); // Code fires
return KV.of(parts[0], Integer.parseInt(parts[1]));
})
)
.apply(GroupByKey.create())
.apply(ParDo.of(new Compute()));
p.run();
}
Background
We have a pipeline that starts by receiving messages from PubSub, each with the name of a file. These files are exploded to line level, parsed to JSON object nodes and then sent to an external decoding service (which decodes some encoded data). Object nodes are eventually converted to Table Rows and written to Big Query.
It appeared that Dataflow was not acknowledging the PubSub messages until they arrived at the decoding service. The decoding service is slow, resulting in a backlog when many message are sent at once. This means that lines associated with a PubSub message can take some time to arrive at the decoding service. As a result, PubSub was receiving no acknowledgement and resending the message. My first attempt to remedy this was adding an attribute to each PubSub messages that is passed to the Reader using withAttributeId(). However, on testing, this only prevented duplicates that arrived close together.
My second attempt was to add a fusion breaker (example) after the PubSub read. This simply performs a needless GroupByKey and then ungroups, the idea being that the GroupByKey forces Dataflow to acknowledge the PubSub message.
The Problem
The fusion breaker discussed above works in that it prevents PubSub from resending messages, but I am finding that this GroupByKey outputs more elements than it receives: See image.
To try and diagnose this I have removed parts of the pipeline to get a simple pipeline that still exhibits this behavior. The behavior remains even when
PubSub is replaced by some dummy transforms that send out a fixed list of messages with a slight delay between each one.
The Writing transforms are removed.
All Side Inputs/Outputs are removed.
The behavior I have observed is:
Some number of the received messages pass straight through the GroupByKey.
After a certain point, messages are 'held' by the GroupByKey (presumably due to the backlog after the GroupByKey).
These messages eventually exit the GroupByKey (in groups of size one).
After a short delay (about 3 minutes), the same messages exit the GroupByKey again (still in groups of size one). This may happen several times (I suspect it is proportional to the time they spend waiting to enter the GroupByKey).
Example job id is 2017-10-11_03_50_42-6097948956276262224. I have not run the beam on any other runner.
The Fusion Breaker is below:
#Slf4j
public class FusionBreaker<T> extends PTransform<PCollection<T>, PCollection<T>> {
#Override
public PCollection<T> expand(PCollection<T> input) {
return group(window(input.apply(ParDo.of(new PassthroughLogger<>(PassthroughLogger.Level.Info, "Fusion break in")))))
.apply("Getting iterables after breaking fusion", Values.create())
.apply("Flattening iterables after breaking fusion", Flatten.iterables())
.apply(ParDo.of(new PassthroughLogger<>(PassthroughLogger.Level.Info, "Fusion break out")));
}
private PCollection<T> window(PCollection<T> input) {
return input.apply("Windowing before breaking fusion", Window.<T>into(new GlobalWindows())
.triggering(Repeatedly.forever(AfterPane.elementCountAtLeast(1)))
.discardingFiredPanes());
}
private PCollection<KV<Integer, Iterable<T>>> group(PCollection<T> input) {
return input.apply("Keying with random number", ParDo.of(new RandomKeyFn<>()))
.apply("Grouping by key to break fusion", GroupByKey.create());
}
private static class RandomKeyFn<T> extends DoFn<T, KV<Integer, T>> {
private Random random;
#Setup
public void setup() {
random = new Random();
}
#ProcessElement
public void processElement(ProcessContext context) {
context.output(KV.of(random.nextInt(), context.element()));
}
}
}
The PassthroughLoggers simply log the elements passing through (I use these to confirm that elements are indeed repeated, rather than there being an issue with the counts).
I suspect this is something to do with windows/triggers, but my understanding is that elements should never be repeated when .discardingFiredPanes() is used - regardless of the windowing setup. I have also tried FixedWindows with no success.
First, the Reshuffle transform is equivalent to your Fusion Breaker, but has some additional performance improvements that should make it preferable.
Second, both counters and logging may see an element multiple times if it is retried. As described in the Beam Execution Model, an element at a step may be retried if anything that is fused into it is retried.
Have you actually observed duplicates in what is written as the output of the pipeline?
I have a Pub/Sub topic + subscription and want to consume and aggregate the unbounded data from the subscription in a Dataflow. I use a fixed window and write the aggregates to BigQuery.
Reading and writing (without windowing and aggregation) works fine. But when I pipe the data into a fixed window (to count the elements in each window) the window is never triggered. And thus the aggregates are not written.
Here is my word publisher (it uses kinglear.txt from the examples as input file):
public static class AddCurrentTimestampFn extends DoFn<String, String> {
#ProcessElement public void processElement(ProcessContext c) {
c.outputWithTimestamp(c.element(), new Instant(System.currentTimeMillis()));
}
}
public static class ExtractWordsFn extends DoFn<String, String> {
#ProcessElement public void processElement(ProcessContext c) {
String[] words = c.element().split("[^a-zA-Z']+");
for (String word:words){ if(!word.isEmpty()){ c.output(word); }}
}
}
// main:
Pipeline p = Pipeline.create(o); // 'o' are the pipeline options
p.apply("ReadLines", TextIO.Read.from(o.getInputFile()))
.apply("Lines2Words", ParDo.of(new ExtractWordsFn()))
.apply("AddTimestampFn", ParDo.of(new AddCurrentTimestampFn()))
.apply("WriteTopic", PubsubIO.Write.topic(o.getTopic()));
p.run();
Here is my windowed word counter:
Pipeline p = Pipeline.create(o); // 'o' are the pipeline options
BigQueryIO.Write.Bound tablePipe = BigQueryIO.Write.to(o.getTable(o))
.withSchema(o.getSchema())
.withCreateDisposition(BigQueryIO.Write.CreateDisposition.CREATE_IF_NEEDED)
.withWriteDisposition(BigQueryIO.Write.WriteDisposition.WRITE_APPEND);
Window.Bound<String> w = Window
.<String>into(FixedWindows.of(Duration.standardSeconds(1)));
p.apply("ReadTopic", PubsubIO.Read.subscription(o.getSubscription()))
.apply("FixedWindow", w)
.apply("CountWords", Count.<String>perElement())
.apply("CreateRows", ParDo.of(new WordCountToRowFn()))
.apply("WriteRows", tablePipe);
p.run();
The above subscriber will not work, since the window does not seem to trigger using the default trigger. However, if I manually define a trigger the code works and the counts are written to BigQuery.
Window.Bound<String> w = Window.<String>into(FixedWindows.of(Duration.standardSeconds(1)))
.triggering(AfterProcessingTime
.pastFirstElementInPane()
.plusDelayOf(Duration.standardSeconds(1)))
.withAllowedLateness(Duration.ZERO)
.discardingFiredPanes();
I like to avoid specifying custom triggers if possible.
Questions:
Why does my solution not work with Dataflow's default trigger?
How do I have to change my publisher or subscriber to trigger windows using the default trigger?
How are you determining the trigger never fires?
Your PubSubIO.Write and PubSubIO.Read transforms should both specify a timestamp label using withTimestampLabel, otherwise the timestamps you've added will not be written to PubSub and the publish times will be used.
Either way, the input watermark of the pipeline will be derived from the timestamps of the elements waiting in PubSub. Once all inputs have been processed, it will stay back for a few minutes (in case there was a delay in the publisher) before advancing to real time.
What you are likely seeing is that all the elements are published in the same ~1 second window (since the input file is pretty small). These are all read and processed relatively quickly, but the 1-second window they are put in will not trigger until after the input watermark has advanced, indicating that all data in that 1-second window has been consumed.
This won't happen until several minutes, which may make it look like the trigger isn't working. The trigger you wrote fired after 1 second of processing time, which would fire much earlier, but there is no guarantee all the data has been processed.
Steps to get better behavior from the default trigger:
Use withTimestampLabel on both the write and read pubsub steps.
Have the publisher spread the timestamps out further (eg., run for several minutes and spread the timestamps out across that range)