We are streaming messages to a Kafka topic at a rate of a few hundred per second. Each message has a timestamp and a payload. Ultimately, we would like aggregate one hour worth of data - based on the timestamp of the message - into parquet files and upload them to a cheap remote storage (object-store).
A naive approach would be to have the consumer simply read the messages from the topic and do the aggregation/roll-up in memory, and once there is one hour worth of data, generate and upload the parquet file.
However, in case the consumer crashes or needs to be restarted, we would lose all data since the beginning of the current hour - if we use enable.auto.commit=true or enable.auto.commit=false and manually commit after a batch of messages.
A simple solution for the Consumer could be to keep reading until one hour worth of data is in memory, do the parquet file generation (and upload it), and only then call commitAsync() or commitSync() (using enable.auto.commit=false and use an external store to keep track of the offsets).
But this would lead to millions of messages not being committed for at least one hour. I am wondering if Kafka does even allow to "delay" the commit of messages for so many messages / so long time (I seem to remember to have read about this somewhere but for the life of me I cannot find it again).
Actual questions:
a) is there a limit to the number of messages (or duration) not being committed before Kafka possibly considers the Consumer to be broken or stops giving additional messages to the consumer? this seems counter-intuitive though, since what would be the purpose of enable.auto.commit=false and managing the offsets in the Consumer (with e.g. the help of an external database).
b) in terms of robustness/redundancy and scalability, it would be great to have more than one Consumer in the consumer group; if I understand correctly, it is never possible to have more than one Consumer per partition. If we then run more than one Consumer and configure multiple partitions per topic we cannot do this kind of aggregation/roll-up, since now messages will be distributed across Consumers. The only way to work-around this issue would be to have an additional (external) temporary storage for all those messages belonging to such one-hour group, correct?
You can configure Kafka Streams with a TimestampExtractor to aggregate data into different types of time-windows
into parquet files and upload them to a cheap remote storage (object-store).
Kafka Connect S3 sink, or Pinterest Secor tool, already do this
Related
Having a Kafka Streams application, that performs windowing(using original event time, not wallclock time) via Stream joins of e.g. 1 day.
If bringing up this topology, and reprocessing the data from the start (as in a lambda-style architecture), will this window keep that old data there? da
For example: if today is 2022-01-09, and I'm receiving data from 2021-03-01, will this old data enter the table, or will it be rejected from the start?
In that case - what strategies can be done to reprocess this data?
UPDATE Using Kafka Streams 2.5.0
Updated Answer to OP Kafka Streams version 2.5:
When using event time, Kafka Streams will behave independent of the wallclock time, as long as no events contain the wallclock time. You should not have configured a WallclockTimestampExtractor as your timestamp extractor.
Kafka Streams will assign you input topic partitions to stream tasks, that will consume the partitions one event at a time. On any given topic, at most one partition will be assigned to a stream task. Time-windowed aggregations are carried out for each stream task separately. Kafka Streams uses an internal timestamp called "observedStreamTime" for each aggregation to keep track of the maximum timestamp seen so far. Incoming records are checked for their timestamp in comparison to the observedStreamTime. If they are older than the retention + grace period of the configured time window store, they will be dropped. Otherwise, they will be aggregated according to the configuration. The implementation can be found at https://github.com/apache/kafka/blob/d5b53ad132d1c1bfcd563ce5015884b6da831777/streams/src/main/java/org/apache/kafka/streams/kstream/internals/KStreamWindowAggregate.java#L108-L175
This processing will always yield the same result, if the Kafka Streams application is reset. It is independent on the execution time of the processing. If events are dropped, the corresponding metrics are changed.
There is one caveat with this approach, when multiple topics are consumed. The observedStreamTime will reflect the highest timestamp of all partitions read by the stream task. If you have two topics (maybe because you want to join them) and one contains considerably younger data than the other (maybe because the latter received no new data), the observedStreamTime will be dominated by the younger topic. Events of the older topic might be dropped, if the time window configuration does not have enough retention or grace periods. See the JavaDoc of TimeWindows on the configuration options: https://github.com/apache/kafka/blob/d5b53ad132d1c1bfcd563ce5015884b6da831777/streams/src/main/java/org/apache/kafka/streams/kstream/TimeWindows.java
In your example the old data will be accepted, as long as the stream time has not progress too far. Reprocessing the whole data set should work, since it will linearly progress through your topic. If the old data is aggregated in a time-window with exceeding the window size + grace period, Kafka Streams will reject the record. In that case Kafka Streams will also issue an error message and adjust its metrics accordingly. So this behaviour should be easy to pick up.
I suggest to try out this reprocessing if feasible and watch the logs and metrics.
I have a usecase where I want to have thousands of producers writing messages which will be consumed by thousands of corresponding consumers. Each producer's message is meant for exactly one consumer.
Going through the core concepts here and here: it seems like each consumer-producer pair should have its own topic. Is this correct understanding? I also looked into consumer groups but it seems they are more for parallellizing consumption.
Right now I have multiple producer-consumer pairs sharing very few topics, but because of that (i think) I am having to read a lot of messages in the consumer and filter them out for the specific producer's messages by the key. As my system scales this might take a lot of time. Also in the event I have to delete the checkpoint this will be even more problematic as it starts reading from the very beginning.
Is creating thousands of topics the solution for this? Or is there any other way to use concepts like partitions, consumer groups etc? Both producers and consumers are spark streaming/batch applications. Thanks.
Each producer's message is meant for exactly one consumer
Assuming you commit the offsets, and don't allow retries, this is the expected behavior of all Kafka consumers (or rather, consumer groups)
seems like each consumer-producer pair should have its own topic
Not really. As you said, you have many-to-many relationship of clients. You do not need to have a known pair ahead of time; a producer could send data with no expected consumer, then any consumer application(s) in the future should be able to subscribe to that topic for the data they are interested in.
sharing very few topics, but because of that (i think) I am having to read a lot of messages in the consumer and filter them out for the specific producer's messages by the key. As my system scales this might take a lot of time
The consumption would take linearly more time on a higher production rate, yes, and partitions are the way to solve for that. Beyond that, you need faster network and processing. You still need to consume and deserialize in order to filter, so the filter is not the bottleneck here.
Is creating thousands of topics the solution for this?
Ultimately depends on your data, but I'm guessing not.
Is creating thousands of topics the solution for this? Or is there any
other way to use concepts like partitions, consumer groups etc? Both
producers and consumers are spark streaming/batch applications.
What's the reason you want to have thousands of consumers? or want to have a 1 to 1 explicit relationship? As mentioned earlier, only one consumer within a consumer group will process a message. This is normal.
If however you are trying to make your record processing extremely concurrent, instead of using very high partition counts or very large consumer groups, should use something like Parallel Consumer (PC).
By using PC, you can processing all your keys in parallel, regardless of how long it takes to process, and you can be as concurrent as you wish .
PC directly solves for this, by sub partitioning the input partitions by key and processing each key in parallel.
It also tracks per record acknowledgement. Check out Parallel Consumer on GitHub (it's open source BTW, and I'm the author).
We have a Kafka producer that produces keyed messages in a very high frequency to topics whose retention time = 10 hours. These messages are real-time updates and the used key is the ID of the element whose value has changed. So the topic is acting as a changelog and will have many duplicate keys.
Now, what we're trying to achieve is that when a Kafka consumer launches, regardless of the last known state (new consumer, crashed, restart, etc..), it will somehow construct a table with the latest values of all the keys in a topic, and then keeps listening for new updates as normal, keeping the minimum load on Kafka server and letting the consumer do most of the job. We tried many ways and none of them seems the best.
What we tried:
1 changelog topic + 1 compact topic:
The producer sends the same message to both topics wrapped in a transaction to assure successful send.
Consumer launches and requests the latest offset of the changelog topic.
Consumes the compacted topic from beginning to construct the table.
Continues consuming the changelog since the requested offset.
Cons:
Having duplicates in compacted topic is a very high possibility even with setting the log compaction frequency the highest possible.
x2 number of topics on Kakfa server.
KSQL:
With KSQL we either have to rewrite a KTable as a topic so that consumer can see it (Extra topics), or we will need consumers to execute KSQL SELECT using to KSQL Rest Server and query the table (Not as fast and performant as Kafka APIs).
Kafka Consumer API:
Consumer starts and consumes the topic from beginning. This worked perfectly, but the consumer has to consume the 10 hours change log to construct the last values table.
Kafka Streams:
By using KTables as following:
KTable<Integer, MarketData> tableFromTopic = streamsBuilder.table("topic_name", Consumed.with(Serdes.Integer(), customSerde));
KTable<Integer, MarketData> filteredTable = tableFromTopic.filter((key, value) -> keys.contains(value.getRiskFactorId()));
Kafka Streams will create 1 topic on Kafka server per KTable (named {consumer_app_id}-{topic_name}-STATE-STORE-0000000000-changelog), which will result in a huge number of topics since we a big number of consumers.
From what we have tried, it looks like we need to either increase the server load, or the consumer launch time. Isn't there a "perfect" way to achieve what we're trying to do?
Thanks in advance.
By using KTables, Kafka Streams will create 1 topic on Kafka server per KTable, which will result in a huge number of topics since we a big number of consumers.
If you are just reading an existing topic into a KTable (via StreamsBuilder#table()), then no extra topics are being created by Kafka Streams. Same for KSQL.
It would help if you could clarify what exactly you want to do with the KTable(s). Apparently you are doing something that does result in additional topics being created?
1 changelog topic + 1 compact topic:
Why were you thinking about having two separate topics? Normally, changelog topics should always be compacted. And given your use case description, I don't see a reason why it should not be:
Now, what we're trying to achieve is that when a Kafka consumer launches, regardless of the last known state (new consumer, crashed, restart, etc..), it will somehow construct a table with the latest values of all the keys in a topic, and then keeps listening for new updates as normal [...]
Hence compaction would be very useful for your use case. It would also prevent this problem you described:
Consumer starts and consumes the topic from beginning. This worked perfectly, but the consumer has to consume the 10 hours change log to construct the last values table.
Note that, to reconstruct the latest table values, all three of Kafka Streams, KSQL, and the Kafka Consumer must read the table's underlying topic completely (from beginning to end). If that topic is NOT compacted, this might indeed take a long time depending on the data volume, topic retention settings, etc.
From what we have tried, it looks like we need to either increase the server load, or the consumer launch time. Isn't there a "perfect" way to achieve what we're trying to do?
Without knowing more about your use case, particularly what you want to do with the KTable(s) once they are populated, my answer would be:
Make sure the "changelog topic" is also compacted.
Try KSQL first. If this doesn't satisfy your needs, try Kafka Streams. If this doesn't satisfy your needs, try the Kafka Consumer.
For example, I wouldn't use the Kafka Consumer if it is supposed to do any stateful processing with the "table" data, because the Kafka Consumer lacks built-in functionality for fault-tolerant stateful processing.
Consumer starts and consumes the topic from beginning. This worked
perfectly, but the consumer has to consume the 10 hours change log to
construct the last values table.
During the first time your application starts up, what you said is correct.
To avoid this during every restart, store the key-value data in a file.
For example, you might want to use a persistent map (like MapDB).
Since you give the consumer group.id and you commit the offset either periodically or after each record is stored in the map, the next time your application restarts it will read it from the last comitted offset for that group.id.
So the problem of taking a lot of time occurs only initially (during first time). So long as you have the file, you don't need to consume from beginning.
In case, if the file is not there or is deleted, just seekToBeginning in the KafkaConsumer and build it again.
Somewhere, you need to store this key-values for retrieval and why cannot it be a persistent store?
In case if you want to use Kafka streams for whatever reason, then an alternative (not as simple as the above) is to use a persistent backed store.
For example, a persistent global store.
streamsBuilder.addGlobalStore(Stores.keyValueStoreBuilder(Stores.persistentKeyValueStore(topic), keySerde, valueSerde), topic, Consumed.with(keySerde, valueSerde), this::updateValue);
P.S: There will be a file called .checkpoint in the directory which stores the offsets. In case if the topic is deleted in the middle you get OffsetOutOfRangeException. You may want to avoid this, perhaps by using UncaughtExceptionHandler
Refer to https://stackoverflow.com/a/57301986/2534090 for more.
Finally,
It is better to use Consumer with persistent file rather than Streams for this, because of simplicity it offers.
I have been trying to implement a queuing mechanism using kafka where I want to ensure that duplicate records are not inserted into topic created.
I found that iteration is possible in consumer. Is there any way by which we can do this in producer thread as well?
This is known as exactly-once processing.
You might be interested in the first part of Kafka FAQ that describes some approaches on how to avoid duplication on data production (i.e. on producer side):
Exactly once semantics has two parts: avoiding duplication during data
production and avoiding duplicates during data consumption.
There are two approaches to getting exactly once semantics during data
production:
Use a single-writer per partition and every time you get a network
error check the last message in that partition to see if your last
write succeeded
Include a primary key (UUID or something) in the
message and deduplicate on the consumer.
If you do one of these things, the log that Kafka hosts will be
duplicate-free. However, reading without duplicates depends on some
co-operation from the consumer too. If the consumer is periodically
checkpointing its position then if it fails and restarts it will
restart from the checkpointed position. Thus if the data output and
the checkpoint are not written atomically it will be possible to get
duplicates here as well. This problem is particular to your storage
system. For example, if you are using a database you could commit
these together in a transaction. The HDFS loader Camus that LinkedIn
wrote does something like this for Hadoop loads. The other alternative
that doesn't require a transaction is to store the offset with the
data loaded and deduplicate using the topic/partition/offset
combination.
I think there are two improvements that would make this a lot easier:
Producer idempotence could be done automatically and much more cheaply
by optionally integrating support for this on the server.
The existing
high-level consumer doesn't expose a lot of the more fine grained
control of offsets (e.g. to reset your position). We will be working
on that soon
I have a Kafka consumer project which consumes data from a specific Kafka topic. The 90% of the records are processed as soon as I got them but I have the delay processing some of the records (10%).
This these records need to be delayed, I can't commit the records so it may cause Kafka to reassign the partitions to new nodes. In order to avoid that, I can read the same topic twice and delay the fetching data part in the second consumer but it requires deserialization twice so comes with an overhead.
Is it possible the read records using single consumer but have two separate commits with Kafka consumers? It will be basically similar to having two different consumers in terms of commit, consumer.poll will be called from a single consumer but there will be two consumer.commitSync for each batch. I will help me to avoid extra deserialization and also the network cost.
Below mentioned are the things you can do to achieve the above-mentioned task.
Create a pipe Line having two topics(T1, T2) push all the messages (90%) in topic T1 and rest all the messages 10% in topic T2.
Make your Kafka consumer configurable i.e. you can easily pass polling interval, batchSize, and batch timeout whenever you are starting your consumer.
Find a logic/ or if your second topic consumption is time-based then schedule the cron which will start and stop your consumer topic T2 when it is required.
Regarding consumer Groups, you can place both of your topics in the same group or indifferent. It's completely your choice.
By this way you will be keeping the topics clean.and each and every time you need to process the messages you can do it easily by setting up the pipeline just for once.