I want to read data from Bigquery periodically in Beam, and the test codes as below
pipeline.apply("Generate Sequence",
GenerateSequence.from(0).withRate(1, Duration.standardMinutes(2)))
.apply(Window.into(FixedWindows.of(Duration.standardMinutes(2))))
.apply("Read from BQ", new ReadBQ())
.apply("Convert Row",
MapElements.into(TypeDescriptor.of(MyData.class)).via(MyData::fromTableRow))
.apply("Map TableRow", ParDo.of(new MapTableRowV1()))
;
static class ReadBQ extends PTransform<PCollection<Long>, PCollection<TableRow>> {
#Override
public PCollection<TableRow> expand(PCollection<Long> input) {
BigQueryIO.TypedRead<TableRow> rows = BigQueryIO.readTableRows()
.fromQuery("select * from project.dataset.table limit 10")
.usingStandardSql();
return rows.expand(input.getPipeline().begin());
}
}
static class MapTableRowV1 extends DoFn<AdUnitECPM, Void> {
#ProcessElement
public void processElement(ProcessContext pc) {
LOG.info("String of mydata is " + pc.element().toString());
}
}
Since BigQueryIO.TypedRead is related to PBegin, one trick is done in ReadBQ through rows.expand(input.getPipeline().begin()). However, this job does NOT run every two minutes. How to read data from bigquery periodically?
Look at using Looping Timers. That provides the right pattern.
As written your code would only fire once after sequence is built. For fixed windows you would need an input value coming into the Window for it to trigger. For example, have the pipeline read a Pub/Sub input and then have an agent push events every 2 minutes into the topic/sub.
I am going to assume that you are running in streaming mode here; however, another way to do this would be to use a batch job and then run it every 2 mins from Composer. Reason being if your job is idle for effectively 90 secs (2 min - processing time) your streaming job wasting some resources.
One other note: Look at thinning down you column selection in your BigQuery SQL (to save time and money). Look at using some filter on your SQL to pick up a partition or cluster. Look at using #timestamp filter to only scan records that are new in last N. This could give you better control over how you deal with latency and variability at the db level.
As you have mentioned in the question, BigQueryIO read transforms start with PBegin, which puts it at the start of the Graph. In order to achieve what you are looking for, you will need to make use of the BigQuery client libraries directly within a DoFn.
For an example of this have a look at this
transform
Using a normal DoFn for this will be ok for small amounts of data, but for a large amount of data, you will want to look at implementing that logic in a SDF.
Related
I have a piece of code that is essentially executing the following with Infinispan in embedded mode, using version 13.0.0 of the -core and -clustered-lock modules:
#Inject
lateinit var lockManager: ClusteredLockManager
private fun getLock(lockName: String): ClusteredLock {
lockManager.defineLock(lockName)
return lockManager.get(lockName)
}
fun createSession(sessionId: String) {
tryLockCounter.increment()
logger.debugf("Trying to start session %s. trying to acquire lock", sessionId)
Future.fromCompletionStage(getLock(sessionId).lock()).map {
acquiredLockCounter.increment()
logger.debugf("Starting session %s. Got lock", sessionId)
}.onFailure {
logger.errorf(it, "Failed to start session %s", sessionId)
}
}
I take this piece of code and deploy it to kubernetes. I then run it in six pods distributed over six nodes in the same region. The code exposes createSession with random Guids through an API. This API is called and creates sessions in chunks of 500, using a k8s service in front of the pods which means the load gets balanced over the pods. I notice that the execution time to acquire a lock grows linearly with the amount of sessions. In the beginning it's around 10ms, when there's about 20_000 sessions it takes about 100ms and the trend continues in a stable fashion.
I then take the same code and run it, but this time with twelve pods on twelve nodes. To my surprise I see that the performance characteristics are almost identical to when I had six pods. I've been digging in to the code but still haven't figured out why this is, I'm wondering if there's a good reason why infinispan here doesn't seem to perform better with more nodes?
For completeness the configuration of the locks are as follows:
val global = GlobalConfigurationBuilder.defaultClusteredBuilder()
global.addModule(ClusteredLockManagerConfigurationBuilder::class.java)
.reliability(Reliability.AVAILABLE)
.numOwner(1)
and looking at the code the clustered locks is using DIST_SYNC which should spread out the load of the cache onto the different nodes.
UPDATE:
The two counters in the code above are simply micrometer counters. It is through them and prometheus that I can see how the lock creation starts to slow down.
It's correctly observed that there's one lock created per session id, this is per design what we'd like. Our use case is that we want to ensure that a session is running in at least one place. Without going to deep into detail this can be achieved by ensuring that we at least have two pods that are trying to acquire the same lock. The Infinispan library is great in that it tells us directly when the lock holder dies without any additional extra chattiness between pods, which means that we have a "cheap" way of ensuring that execution of the session continues when one pod is removed.
After digging deeper into the code I found the following in CacheNotifierImpl in the core library:
private CompletionStage<Void> doNotifyModified(K key, V value, Metadata metadata, V previousValue,
Metadata previousMetadata, boolean pre, InvocationContext ctx, FlagAffectedCommand command) {
if (clusteringDependentLogic.running().commitType(command, ctx, extractSegment(command, key), false).isLocal()
&& (command == null || !command.hasAnyFlag(FlagBitSets.PUT_FOR_STATE_TRANSFER))) {
EventImpl<K, V> e = EventImpl.createEvent(cache.wired(), CACHE_ENTRY_MODIFIED);
boolean isLocalNodePrimaryOwner = isLocalNodePrimaryOwner(key);
Object batchIdentifier = ctx.isInTxScope() ? null : Thread.currentThread();
try {
AggregateCompletionStage<Void> aggregateCompletionStage = null;
for (CacheEntryListenerInvocation<K, V> listener : cacheEntryModifiedListeners) {
// Need a wrapper per invocation since converter could modify the entry in it
configureEvent(listener, e, key, value, metadata, pre, ctx, command, previousValue, previousMetadata);
aggregateCompletionStage = composeStageIfNeeded(aggregateCompletionStage,
listener.invoke(new EventWrapper<>(key, e), isLocalNodePrimaryOwner));
}
The lock library uses a clustered Listener on the entry modified event, and this one uses a filter to only notify when the key for the lock is modified. It seems to me the core library still has to check this condition on every registered listener, which of course becomes a very big list as the number of sessions grow. I suspect this to be the reason and if it is it would be really really awesome if the core library supported a kind of key filter so that it could use a hashmap for these listeners instead of going through a whole list with all listeners.
I believe you are creating a clustered lock per session id. Is this what you need ? what is the acquiredLockCounter? We are about to deprecate the "lock" method in favour of "tryLock" with timeout since the lock method will block forever if the clustered lock is never acquired. Do you ever unlock the clustered lock in another piece of code? If you shared a complete reproducer of the code will be very helpful for us. Thanks!
I have a simple Flink application, which sums up the events with the same id and timestamp within the last minute:
DataStream<String> input = env
.addSource(consumerProps)
.uid("app");
DataStream<Pixel> pixels = input.map(record -> mapper.readValue(record, Pixel.class));
pixels
.keyBy("id", "timestampRoundedToMinutes")
.timeWindow(Time.minutes(1))
.sum("constant")
.addSink(dynamoDBSink);
env.execute(jobName);
I am trying to test this application with the recommended approach in documentation. I also have looked at this stackoverflow question, but adding the sink hadn't helped.
I do have a #ClassRule as recommended in my test class. The function looks like this:
StreamExecutionEnvironment env=StreamExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(2);
CollectSink.values.clear();
Pixel testPixel1 = Pixel.builder().id(1).timestampRoundedToMinutes("202002261219").constant(1).build();
Pixel testPixel2 = Pixel.builder().id(2).timestampRoundedToMinutes("202002261220").constant(1).build();
Pixel testPixel3 = Pixel.builder().id(1).timestampRoundedToMinutes("202002261219").constant(1).build();
Pixel testPixel4 = Pixel.builder().id(3).timestampRoundedToMinutes("202002261220").constant(1).build();
env.fromElements(testPixel1, testPixel2, testPixel3, testPixel4)
.keyBy("id","timestampRoundedToMinutes")
.timeWindow(Time.minutes(1))
.sum("constant")
.addSink(new CollectSink());
JobExecutionResult result = env.execute("AggregationTest");
assertNotEquals(0, CollectSink.values.size());
CollectSink is copied from documentation.
What am I doing wrong? Is there also a simple way to test the application with embedded kafka?
Thanks!
The reason why your test is failing is because the window is never triggered. The job runs to completion before the window can reach the end of its allotted time.
The reason for this has to do with the way you are working with time. By specifying
.keyBy("id","timestampRoundedToMinutes")
you are arranging for all the events for the same id and with timestamps within the same minute to be in the same window. But because you are using processing time windowing (rather than event time windowing), your windows won't close until the time of day when the test is running crosses over the boundary from one minute to the next. With only four events to process, your job is highly unlikely to run long enough for this to happen.
What you should do instead is something more like this: set the time characteristic to event time, and provide a timestamp extractor and watermark assigner. Note that by doing this, there's no need to key by the timestamp rounded to minute boundaries -- that's part of what event time windows do anyway.
public static void main(String[] args) throws Exception {
...
env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime);
env.fromElements(testPixel1, testPixel2, testPixel3, testPixel4)
.assignTimestampsAndWatermarks(new TimestampsAndWatermarks())
.keyBy("id")
.timeWindow(Time.minutes(1))
.sum("constant")
.addSink(new CollectSink());
env.execute();
}
private static class TimestampsAndWatermarks extends BoundedOutOfOrdernessTimestampExtractor<Event> {
public TimestampsAndWatermarks() {
super(/* delay to handle out-of-orderness */);
}
#Override
public long extractTimestamp(Event event) {
return event.timestamp;
}
}
See the documentation and the tutorials for more about event time, watermarks, and windowing.
We are trying to write to Big Query using Apache Beam and avro.
The following seems to work ok:-
p.apply("Input", AvroIO.read(DataStructure.class).from("AvroSampleFile.avro"))
.apply("Transform", ParDo.of(new CustomTransformFunction()))
.apply("Load", BigQueryIO.writeTableRows().to(table).withSchema(schema));
We then tried to use it in the following manner to get data from the Google Pub/Sub
p.begin()
.apply("Input", PubsubIO.readAvros(DataStructure.class).fromTopic("topicName"))
.apply("Transform", ParDo.of(new CustomTransformFunction()))
.apply("Write", BigQueryIO.writeTableRows()
.to(table)
.withSchema(schema)
.withTimePartitioning(timePartitioning)
.withCreateDisposition(BigQueryIO.Write.CreateDisposition.CREATE_IF_NEEDED)
.withWriteDisposition(BigQueryIO.Write.WriteDisposition.WRITE_APPEND));
p.run().waitUntilFinish();
When we do this it always pushes it to the buffer and Big Query seems to take a long time to read from the buffer. Can anyone tell me why the above won't write the records directly to the Big Query tables?
UPDATE:-
It looks like I need add the following settings but this throws an java.lang.IllegalArgumentException.
.withMethod(Method.FILE_LOADS)
.withTriggeringFrequency(org.joda.time.Duration.standardMinutes(2))
The answer is you need to include "withNumFileShards" like so (Can be 1 to 1000).
p.begin()
.apply("Input", PubsubIO.readAvros(DataStructure.class).fromTopic("topicName"))
.apply("Transform", ParDo.of(new CustomTransformFunction()))
.apply("Write", BigQueryIO.writeTableRows()
.to(table)
.withSchema(schema)
.withTimePartitioning(timePartitioning)
.withMethod(Method.FILE_LOADS)
.withTriggeringFrequency(org.joda.time.Duration.standardMinutes(2))
.withNumFileShards(1000)
.withCreateDisposition(BigQueryIO.Write.CreateDisposition.CREATE_IF_NEEDED)
.withWriteDisposition(BigQueryIO.Write.WriteDisposition.WRITE_APPEND));
p.run().waitUntilFinish();
I can't find this documented anywhere to say that withNumFileShards is mandatory however there is a Jira ticket for this which I found after the fix.
https://issues.apache.org/jira/browse/BEAM-3198
I have a Beam job running on Google Cloud DataFlow that reads data from BigQuery. When I run the job it takes minutes for the job to start reading data from the (tiny) table. It turns out the dataflow job sends of a BigQuery job which runs in BATCH mode and not in INTERACTIVE mode. How can I switch this to run immediately in Apache Beam? I couldn't find a method in the API to change the priority.
Maybe a Googler will correct me, but no, you cannot change this from BATCH to INTERACTIVE because it's not exposed by Beam's API.
From org.apache.beam.sdk.io.gcp.bigquery.BigQueryIO.java (here):
private void executeQuery(
String executingProject,
String jobId,
TableReference destinationTable,
JobService jobService) throws IOException, InterruptedException {
JobReference jobRef = new JobReference()
.setProjectId(executingProject)
.setJobId(jobId);
JobConfigurationQuery queryConfig = createBasicQueryConfig()
.setAllowLargeResults(true)
.setCreateDisposition("CREATE_IF_NEEDED")
.setDestinationTable(destinationTable)
.setPriority("BATCH") <-- NOT EXPOSED
.setWriteDisposition("WRITE_EMPTY");
jobService.startQueryJob(jobRef, queryConfig);
Job job = jobService.pollJob(jobRef, JOB_POLL_MAX_RETRIES);
if (parseStatus(job) != Status.SUCCEEDED) {
throw new IOException(String.format(
"Query job %s failed, status: %s.", jobId, statusToPrettyString(job.getStatus())));
}
}
If it's really a problem for you that the query is running in BATCH mode, then one workaround could be:
Using the BigQuery API directly, roll your own initial request, and set the priority to INTERACTIVE.
Write the results of step 1 to a temp table
In your Beam pipeline, read the temp table using BigQueryIO.Read.from()
You can configure to run the queries with "Interactive" priority by passing a priority parameter. Check this Github example for details.
Please note that you might be reaching one of the BigQuery limits and quotas as when you use batch, if you ever hit a rate limit, the query will be queued and retried later. As opposed to the interactive ones, when if these limits are hit, the query will fail immediately. This is because BigQuery assumes that an interactive query is something you need run immediately.
I need to design a Redis-driven scalable task scheduling system.
Requirements:
Multiple worker processes.
Many tasks, but long periods of idleness are possible.
Reasonable timing precision.
Minimal resource waste when idle.
Should use synchronous Redis API.
Should work for Redis 2.4 (i.e. no features from upcoming 2.6).
Should not use other means of RPC than Redis.
Pseudo-API: schedule_task(timestamp, task_data). Timestamp is in integer seconds.
Basic idea:
Listen for upcoming tasks on list.
Put tasks to buckets per timestamp.
Sleep until the closest timestamp.
If a new task appears with timestamp less than closest one, wake up.
Process all upcoming tasks with timestamp ≤ now, in batches (assuming
that task execution is fast).
Make sure that concurrent worker wouldn't process same tasks. At the same time, make sure that no tasks are lost if we crash while processing them.
So far I can't figure out how to fit this in Redis primitives...
Any clues?
Note that there is a similar old question: Delayed execution / scheduling with Redis? In this new question I introduce more details (most importantly, many workers). So far I was not able to figure out how to apply old answers here — thus, a new question.
Here's another solution that builds on a couple of others [1]. It uses the redis WATCH command to remove the race condition without using lua in redis 2.6.
The basic scheme is:
Use a redis zset for scheduled tasks and redis queues for ready to run tasks.
Have a dispatcher poll the zset and move tasks that are ready to run into the redis queues. You may want more than 1 dispatcher for redundancy but you probably don't need or want many.
Have as many workers as you want which do blocking pops on the redis queues.
I haven't tested it :-)
The foo job creator would do:
def schedule_task(queue, data, delay_secs):
# This calculation for run_at isn't great- it won't deal well with daylight
# savings changes, leap seconds, and other time anomalies. Improvements
# welcome :-)
run_at = time.time() + delay_secs
# If you're using redis-py's Redis class and not StrictRedis, swap run_at &
# the dict.
redis.zadd(SCHEDULED_ZSET_KEY, run_at, {'queue': queue, 'data': data})
schedule_task('foo_queue', foo_data, 60)
The dispatcher(s) would look like:
while working:
redis.watch(SCHEDULED_ZSET_KEY)
min_score = 0
max_score = time.time()
results = redis.zrangebyscore(
SCHEDULED_ZSET_KEY, min_score, max_score, start=0, num=1, withscores=False)
if results is None or len(results) == 0:
redis.unwatch()
sleep(1)
else: # len(results) == 1
redis.multi()
redis.rpush(results[0]['queue'], results[0]['data'])
redis.zrem(SCHEDULED_ZSET_KEY, results[0])
redis.exec()
The foo worker would look like:
while working:
task_data = redis.blpop('foo_queue', POP_TIMEOUT)
if task_data:
foo(task_data)
[1] This solution is based on not_a_golfer's, one at http://www.saltycrane.com/blog/2011/11/unique-python-redis-based-queue-delay/, and the redis docs for transactions.
You didn't specify the language you're using. You have at least 3 alternatives of doing this without writing a single line of code in Python at least.
Celery has an optional redis broker.
http://celeryproject.org/
resque is an extremely popular redis task queue using redis.
https://github.com/defunkt/resque
RQ is a simple and small redis based queue that aims to "take the good stuff from celery and resque" and be much simpler to work with.
http://python-rq.org/
You can at least look at their design if you can't use them.
But to answer your question - what you want can be done with redis. I've actually written more or less that in the past.
EDIT:
As for modeling what you want on redis, this is what I would do:
queuing a task with a timestamp will be done directly by the client - you put the task in a sorted set with the timestamp as the score and the task as the value (see ZADD).
A central dispatcher wakes every N seconds, checks out the first timestamps on this set, and if there are tasks ready for execution, it pushes the task to a "to be executed NOW" list. This can be done with ZREVRANGEBYSCORE on the "waiting" sorted set, getting all items with timestamp<=now, so you get all the ready items at once. pushing is done by RPUSH.
workers use BLPOP on the "to be executed NOW" list, wake when there is something to work on, and do their thing. This is safe since redis is single threaded, and no 2 workers will ever take the same task.
once finished, the workers put the result back in a response queue, which is checked by the dispatcher or another thread. You can add a "pending" bucket to avoid failures or something like that.
so the code will look something like this (this is just pseudo code):
client:
ZADD "new_tasks" <TIMESTAMP> <TASK_INFO>
dispatcher:
while working:
tasks = ZREVRANGEBYSCORE "new_tasks" <NOW> 0 #this will only take tasks with timestamp lower/equal than now
for task in tasks:
#do the delete and queue as a transaction
MULTI
RPUSH "to_be_executed" task
ZREM "new_tasks" task
EXEC
sleep(1)
I didn't add the response queue handling, but it's more or less like the worker:
worker:
while working:
task = BLPOP "to_be_executed" <TIMEOUT>
if task:
response = work_on_task(task)
RPUSH "results" response
EDit: stateless atomic dispatcher :
while working:
MULTI
ZREVRANGE "new_tasks" 0 1
ZREMRANGEBYRANK "new_tasks" 0 1
task = EXEC
#this is the only risky place - you can solve it by using Lua internall in 2.6
SADD "tmp" task
if task.timestamp <= now:
MULTI
RPUSH "to_be_executed" task
SREM "tmp" task
EXEC
else:
MULTI
ZADD "new_tasks" task.timestamp task
SREM "tmp" task
EXEC
sleep(RESOLUTION)
If you're looking for ready solution on Java. Redisson is right for you. It allows to schedule and execute tasks (with cron-expression support) in distributed way on Redisson nodes using familiar ScheduledExecutorService api and based on Redis queue.
Here is an example. First define a task using java.lang.Runnable interface. Each task can access to Redis instance via injected RedissonClient object.
public class RunnableTask implements Runnable {
#RInject
private RedissonClient redissonClient;
#Override
public void run() throws Exception {
RMap<String, Integer> map = redissonClient.getMap("myMap");
Long result = 0;
for (Integer value : map.values()) {
result += value;
}
redissonClient.getTopic("myMapTopic").publish(result);
}
}
Now it's ready to sumbit it into ScheduledExecutorService:
RScheduledExecutorService executorService = redisson.getExecutorService("myExecutor");
ScheduledFuture<?> future = executorService.schedule(new CallableTask(), 10, 20, TimeUnit.MINUTES);
future.get();
// or cancel it
future.cancel(true);
Examples with cron expressions:
executorService.schedule(new RunnableTask(), CronSchedule.of("10 0/5 * * * ?"));
executorService.schedule(new RunnableTask(), CronSchedule.dailyAtHourAndMinute(10, 5));
executorService.schedule(new RunnableTask(), CronSchedule.weeklyOnDayAndHourAndMinute(12, 4, Calendar.MONDAY, Calendar.FRIDAY));
All tasks are executed on Redisson node.
A combined approach seems plausible:
No new task timestamp may be less than current time (clamp if less). Assuming reliable NTP synch.
All tasks go to bucket-lists at keys, suffixed with task timestamp.
Additionally, all task timestamps go to a dedicated zset (key and score — timestamp itself).
New tasks are accepted from clients via separate Redis list.
Loop: Fetch oldest N expired timestamps via zrangebyscore ... limit.
BLPOP with timeout on new tasks list and lists for fetched timestamps.
If got an old task, process it. If new — add to bucket and zset.
Check if processed buckets are empty. If so — delete list and entrt from zset. Probably do not check very recently expired buckets, to safeguard against time synchronization issues. End loop.
Critique? Comments? Alternatives?
Lua
I made something similar to what's been suggested here, but optimized the sleep duration to be more precise. This solution is good if you have few inserts into the delayed task queue. Here's how I did it with a Lua script:
local laterChannel = KEYS[1]
local nowChannel = KEYS[2]
local currentTime = tonumber(KEYS[3])
local first = redis.call("zrange", laterChannel, 0, 0, "WITHSCORES")
if (#first ~= 2)
then
return "2147483647"
end
local execTime = tonumber(first[2])
local event = first[1]
if (currentTime >= execTime)
then
redis.call("zrem", laterChannel, event)
redis.call("rpush", nowChannel, event)
return "0"
else
return tostring(execTime - currentTime)
end
It uses two "channels". laterChannel is a ZSET and nowChannel is a LIST. Whenever it's time to execute a task, the event is moved from the the ZSET to the LIST. The Lua script with respond with how many MS the dispatcher should sleep until the next poll. If the ZSET is empty, sleep forever. If it's time to execute something, do not sleep(i e poll again immediately). Otherwise, sleep until it's time to execute the next task.
So what if something is added while the dispatcher is sleeping?
This solution works in conjunction with key space events. You basically need to subscribe to the key of laterChannel and whenever there is an add event, you wake up all the dispatcher so they can poll again.
Then you have another dispatcher that uses the blocking left pop on nowChannel. This means:
You can have the dispatcher across multiple instances(i e it's scaling)
The polling is atomic so you won't have any race conditions or double events
The task is executed by any of the instances that are free
There are ways to optimize this even more. For example, instead of returning "0", you fetch the next item from the zset and return the correct amount of time to sleep directly.
Expiration
If you can not use Lua scripts, you can use key space events on expired documents.
Subscribe to the channel and receive the event when Redis evicts it. Then, grab a lock. The first instance to do so will move it to a list(the "execute now" channel). Then you don't have to worry about sleeps and polling. Redis will tell you when it's time to execute something.
execute_later(timestamp, eventId, event) {
SET eventId event EXP timestamp
SET "lock:" + eventId, ""
}
subscribeToEvictions(eventId) {
var deletedCount = DEL eventId
if (deletedCount == 1) {
// move to list
}
}
This however has it own downsides. For example, if you have many nodes, all of them will receive the event and try to get the lock. But I still think it's overall less requests any anything suggested here.