This is our set up, we have one topic and this topic has two subscriptions v1 and v2 with exact identical settings, both are pull subscription with 10 sec ack deadline.
both subscription v1 and v2 goes to separate dedicated dataflow where v2's data flow is more optimized but pretty much doing same thing.
Problem is that every now and then we see below warning messages and backlog starts build up in v2 subscription only and v1 shows little to no backlog.
08:53:56.000 ${MESSAGE_ID} Pubsub processing delay was high at 72 sec.
Dataflow log in v2 shows nothing obvious except above messages. In fact v2 dataflow cpu usage is lower than v1 so I cant make any sense out of this.
Questions:
What causes processing delay and how can I fix it?
Why isn't v1 subscription getting same warnings?
Updated at 2017/01/17
As suggested via #ben it seems that ParDo filtering operation we do right after PubSub read is hitting unexpectedly high latency. But considering getClassroomIds is a simple java list I'm not sure how I can tackle this problem. One question is that is coder we have applied to pubsub lazy? Is unzipping and deserializing we have defined in coder applied when ProcessContext#element() is called?
def processElement(c: DoFn[Entity, Entity]#ProcessContext) = {
val startTime = System.currentTimeMillis()
val entity = c.element()
if (!entity.getClassroomIds.isEmpty) {
c.output(entity)
}
val latencyMs = System.currentTimeMillis() - startTime
if (latencyMs > 1000) {
// We see this warning messages during the load spike
log.warn(s"latency breached 1 second threshold: $latencyMs ms")
}
}
The timing you mention isn't accurately accounting for the time spent in that step. In particular, due to the fusion optimization it reflects all of the ParDos after the filtering operation.
If you have a pipeline which looks like:
ReadFromPubSub -> ParDo(Filter) -> ParDo(Expensive) -> ParDo(Write)
Both Expensive and Write execute on each element coming out of the Filter before the call to c.output returns. This is further a problem because it is fused into the elements coming out of PubSub.
The easiest solution is probably to perform a Reshuffle:
pipeline
.apply(PubSubIO.read(...))
.apply(keyEachElement by their value and Void)
.apply(new Reshuffle<E, Void>())
.apply(move the key back to the element)
// rest of the pipeline
Note that using the Reshuffle instead of a regular GroupByKey has nice properties since it triggers faster than any normally trigger will fire.
Related
We have a Vertex AI model that takes a relatively long time to return a prediction.
When hitting the model endpoint with one instance, things work fine. But batch jobs of size say 1000 instances end up with around 150 504 errors (upstream request timeout). (We actually need to send batches of 65K but I'm troubleshooting with 1000).
I tried increasing the number of replicas assuming that the # of instances handed to the model would be (1000/# of replicas) but that doesn't seem to be the case.
I then read that the default batch size is 64 and so tried decreasing the batch size to 4 like this from the python code that creates the batch job:
model_parameters = dict(batch_size=4)
def run_batch_prediction_job(vertex_config):
aiplatform.init(
project=vertex_config.vertex_project, location=vertex_config.location
)
model = aiplatform.Model(vertex_config.model_resource_name)
model_params = dict(batch_size=4)
batch_params = dict(
job_display_name=vertex_config.job_display_name,
gcs_source=vertex_config.gcs_source,
gcs_destination_prefix=vertex_config.gcs_destination,
machine_type=vertex_config.machine_type,
accelerator_count=vertex_config.accelerator_count,
accelerator_type=vertex_config.accelerator_type,
starting_replica_count=replica_count,
max_replica_count=replica_count,
sync=vertex_config.sync,
model_parameters=model_params
)
batch_prediction_job = model.batch_predict(**batch_params)
batch_prediction_job.wait()
return batch_prediction_job
I've also tried increasing the machine type to n1-high-cpu-16 and that helped somewhat but I'm not sure I understand how batches are sent to replicas?
Is there another way to decrease the number of instances sent to the model?
Or is there a way to increase the timeout?
Is there log output I can use to help figure this out?
Thanks
Answering your follow up question above.
Is that timeout for a single instance request or a batch request. Also, is it in seconds?
This is a timeout for the batch job creation request.
The timeout is in seconds, according to create_batch_prediction_job() timeout refers to rpc timeout. If we trace the code we will end up here and eventually to gapic where timeout is properly described.
timeout (float): The amount of time in seconds to wait for the RPC
to complete. Note that if ``retry`` is used, this timeout
applies to each individual attempt and the overall time it
takes for this method to complete may be longer. If
unspecified, the the default timeout in the client
configuration is used. If ``None``, then the RPC method will
not time out.
What I could suggest is to stick with whatever is working for your prediction model. If ever adding the timeout will improve your model might as well build on it along with your initial solution where you used a machine with a higher spec. You can also try using a machine with higher memory like the n1-highmem-* family.
I'm using Kinesis Data Analytics on Flink to do stream processing.
The usecase that I'm working on is to read records from a single Kinesis stream and after some transformations write to multiple S3 buckets. One source record might end up in multiple S3 buckets. We need to write to multiple buckets since the source record contains a lot of information which needs to be split to multiple S3 buckets.
I tried achieving this using multiple sinks.
private static <T> SinkFunction<T> createS3SinkFromStaticConfig(String path, Class<T> type) {
OutputFileConfig config = OutputFileConfig
.builder()
.withPartSuffix(".snappy.parquet")
.build();
final StreamingFileSink<T> sink = StreamingFileSink
.forBulkFormat(new Path(s3SinkPath + "/" + path), createParquetWriter(type))
.withBucketAssigner(new S3BucketAssigner<T>())
.withOutputFileConfig(config)
.withRollingPolicy(new RollingPolicy<T>(DEFAULT_MAX_PART_SIZE, DEFAULT_ROLLOVER_INTERVAL))
.build();
return sink;
}
public static void main(String[] args) throws Exception {
DataStream<PIData> input = createSourceFromStaticConfig(env)
.map(new JsonToSourceDataMap())
.name("jsonToInputDataTransformation");
input.map(value -> value)
.name("rawData")
.addSink(createS3SinkFromStaticConfig("raw_data", InputData.class))
.name("s3Sink");
input.map(FirstConverter::convertInputData)
.addSink(createS3SinkFromStaticConfig("firstOutput", Output1.class));
input.map(SecondConverter::convertInputData)
.addSink(createS3SinkFromStaticConfig("secondOutput", Output2.class));
input.map(ThirdConverter::convertInputData)
.addSink(createS3SinkFromStaticConfig("thirdOutput", Output3.class));
//and so on; There are around 10 buckets.
}
However, I saw a big performance impact due to this. I saw a big CPU spike due to this (as compared to one with just one sink). The scale that I'm looking at is around 100k records per second.
Other notes:
I'm using bulk format writer since I want to write files in parquet format. I tried increasing the checkpointing interval from 1-minute to 3-minutes assuming writing files to s3 every minute might be causing issues. But this didn't help much.
As I'm new to flink and stream processing, I'm not sure if this much performance impact is expected or is there something I can do better?
Would using a flatmap operator and then having a single sink be better?
When you had a very simple pipeline with a single source and a single sink, something like this:
source -> map -> sink
then the Flink scheduler was able to optimize the execution, and the entire pipeline ran as a sequence of function calls within a single task -- with no serialization or network overhead. Flink 1.12 can apply this operator chaining optimization to more complex topologies -- perhaps including the one you have now with multiple sinks -- but I don't believe this was possible with Flink 1.11 (which is what KDA is currently based on).
I don't see how using a flatmap would make any difference.
You can probably optimize your serialization/deserialization. See https://flink.apache.org/news/2020/04/15/flink-serialization-tuning-vol-1.html.
When I'm inserting rows on BigQuery using writeTableRows, performance is really bad compared to InsertAllRequest. Clearly, something is not setup correctly.
Use case 1: I wrote a Java program to process 'sample' Twitter stream using Twitter4j. When a tweet comes in I write it to BigQuery using this:
insertAllRequestBuilder.addRow(rowContent);
When I run this program from my Mac, it inserts about 1000 rows per minute directly into BigQuery table. I thought I could do better by running a Dataflow job on the cluster.
Use case 2: When a tweet comes in, I write it to a topic of Google's PubSub. I run this from my Mac which sends about 1000 messages every minute.
I wrote a Dataflow job that reads this topic and writes to BigQuery using BigQueryIO.writeTableRows(). I have a 8 machine Dataproc cluster. I started this job on the master node of this cluster with DataflowRunner. It's unbelievably slow! Like 100 rows every 5 minutes or so. Here's a snippet of the relevant code:
statuses.apply("ToBQRow", ParDo.of(new DoFn<Status, TableRow>() {
#ProcessElement
public void processElement(ProcessContext c) throws Exception {
TableRow row = new TableRow();
Status status = c.element();
row.set("Id", status.getId());
row.set("Text", status.getText());
row.set("RetweetCount", status.getRetweetCount());
row.set("FavoriteCount", status.getFavoriteCount());
row.set("Language", status.getLang());
row.set("ReceivedAt", null);
row.set("UserId", status.getUser().getId());
row.set("CountryCode", status.getPlace().getCountryCode());
row.set("Country", status.getPlace().getCountry());
c.output(row);
}
}))
.apply("WriteTableRows", BigQueryIO.writeTableRows().to(tweetsTable)//
.withSchema(schema)
.withMethod(BigQueryIO.Write.Method.FILE_LOADS)
.withTriggeringFrequency(org.joda.time.Duration.standardMinutes(2))
.withNumFileShards(1000)
.withWriteDisposition(BigQueryIO.Write.WriteDisposition.WRITE_APPEND)
.withCreateDisposition(BigQueryIO.Write.CreateDisposition.CREATE_IF_NEEDED));
What am I doing wrong? Should I use a 'SparkRunner'? How do I confirm that it's running on all nodes of my cluster?
With BigQuery you can either:
Stream data in. Low latency, up to 100k rows per second, has a cost.
Batch data in. Way higher latency, incredible throughput, totally free.
That's the difference you are experiencing. If you only want to ingest 1000 rows, batching will be noticeably slower. The same with 10 billion rows will be way faster thru batching, and at no cost.
Dataflow/Bem's BigQueryIO.writeTableRows can either stream or batch data in.
With BigQueryIO.Write.Method.FILE_LOADS the pasted code is choosing batch.
I have an Akka Stream and I want the stream to send messages down stream approximately every second.
I tried two ways to solve this problem, the first way was to make the producer at the start of the stream only send messages once every second when a Continue messages comes into this actor.
// When receive a Continue message in a ActorPublisher
// do work then...
if (totalDemand > 0) {
import scala.concurrent.duration._
context.system.scheduler.scheduleOnce(1 second, self, Continue)
}
This works for a short while then a flood of Continue messages appear in the ActorPublisher actor, I assume (guess but not sure) from downstream via back-pressure requesting messages as the downstream can consume fast but the upstream is not producing at a fast rate. So this method failed.
The other way I tried was via backpressure control, I used a MaxInFlightRequestStrategy on the ActorSubscriber at the end of the stream to limit the number of messages to 1 per second. This works but messages coming in come in at approximately three or so at a time, not just one at a time. It seems the backpressure control doesn't immediately change the rate of messages coming in OR messages were already queued in the stream and waiting to be processed.
So the problem is, how can I have an Akka Stream which can process one message only per second?
I discovered that MaxInFlightRequestStrategy is a valid way to do it but I should set the batch size to 1, its batch size is default 5, which was causing the problem I found. Also its an over-complicated way to solve the problem now that I am looking at the submitted answer here.
You can either put your elements through the throttling flow, which will back pressure a fast source, or you can use combination of tick and zip.
The first solution would be like this:
val veryFastSource =
Source.fromIterator(() => Iterator.continually(Random.nextLong() % 10000))
val throttlingFlow = Flow[Long].throttle(
// how many elements do you allow
elements = 1,
// in what unit of time
per = 1.second,
maximumBurst = 0,
// you can also set this to Enforcing, but then your
// stream will collapse if exceeding the number of elements / s
mode = ThrottleMode.Shaping
)
veryFastSource.via(throttlingFlow).runWith(Sink.foreach(println))
The second solution would be like this:
val veryFastSource =
Source.fromIterator(() => Iterator.continually(Random.nextLong() % 10000))
val tickingSource = Source.tick(1.second, 1.second, 0)
veryFastSource.zip(tickingSource).map(_._1).runWith(Sink.foreach(println))
New to Siddhi CEP. Other than the regular docs on WS02 CEP can someone point to a good tutorial.
Here are our requirements. Point out some clues on the right ways of writing such queries.
Have a single stream of sensor device notification ( IOT application ).
Stream input is via REST-JSON and output is also to be formatted to REST-JSON. ( Hope this is possible on WS02 CEP 3.1)
Kind of execution plan required:
- If device notification reports usage of Sensor 1, then monitor to see if within 5 mins if device notification reports usage of Sensor 2 also. If found then generate output stream reporting composite-activity back on REST-JSON.
- If such composite-activity is not detected during a time slot in morning, afternoon and evening then generate warning-event-stream status on REST-JSON. ( So how to find events which did not occur in time )
- If such composite-activity is not found within some time slots in morning, afternoon and evening then report failure1-event-stream status back on REST-JSON.
This should work day on day, so how will the previous processed data get deleted in WSO2 CEP.
Regards,
Amit
The queries can be as follows (these are draft queries and may require slight modifications to get them running)
To detect sensor 1, and then sensor 2 within 5 minutes (assuming sensorStram has id, value) you can simply use a pattern like following with the 'within' keyword:
from e1=sensorStream[sensorId == '1'] -> e2=sensorStream[sensorId == '2']
select 'composite activity detected' as description, e1.value as sensor1Value, e2.value as sensor2Value
within 5 minutes
insert into compositeActivityStream;
To detect non occurrences (id=1 arrives, but no id=2 within 5 minutes) we can have following two queries:
from sensorStream[sensorId == '1']#window.time(5 minutes)
select *
insert into delayedSensor1Stream for expired-events;
from e1=sensorStream[sensorId == '1'] -> nonOccurringEvent = sensorStream[sensorId == '2'] or delayedEvent=delayedSensor1Stream
select 'id=2 not found' as description, e1.value as id1Value, nonOccurringEvent.sensorId as nonOccurringId
having (not(nonOccurringId instanceof string))
insert into nonOccurrenceStream;
This will detect non-occurrences immediately at the end of 5 minutes after the arrival of id=1 event.
For an explanation of the above logic, have a look at the non occurrence sample of cep 4.0.0 (the syntax is a bit different, but the same idea)
Now since you need to periodically generate a report, we need another query. For convenience i assume you need a report every 6 hours (360 minutes) and use a time batch window here. Alternatively with the new CEP 4.0.0 you can use the 'Cron window' to generate this at specific times which is better for your use case.
from nonOccurrenceStream#window.timeBatch(360 minutes)
select count(id1Value) as nonOccurrenceCount
insert into nonOccurrenceReportsStream for expired-events;
You can use http input/output adaptors and do json mappings with json builders and formatters for this use case.