You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@flink.apache.org by aj <aj...@gmail.com> on 2020/12/04 15:20:22 UTC
Re: Broadcasting control messages to a sink
Hi Jafee,
Can u please help me out with the sample code how you have written the
custom sink and how you using this broadcast pattern to update schema at
run time. It will help me.
On Sat, Oct 17, 2020 at 1:55 PM Piotr Nowojski <pn...@apache.org> wrote:
> Hi Julian,
>
> Glad to hear it worked! And thanks for coming back to us :)
>
> Best,
> Piotrek
>
> sob., 17 paź 2020 o 04:22 Jaffe, Julian <Ju...@activision.com>
> napisał(a):
>
>> Hey Piotr,
>>
>>
>>
>> Thanks for your help! The main thing I was missing was the .broadcast
>> partition operation on a stream (searching for “broadcasting” obviously
>> brought up the broadcast state pattern). This coupled with my
>> misunderstanding of an error in my code as being an error in Flink code
>> resulted in me making this a much harder problem than it needed to be.
>>
>>
>>
>> For anyone who may find this in the future, Piotr’s suggestion is pretty
>> spot-on. I wound up broadcasting (as in the partitioning strategy) my
>> schema stream and connecting it to my event stream. I then processed those
>> using a CoProcessFunction, using the schema messages to update the parsing
>> for the events. I also emitted a side output message when I processed a new
>> schema, using the same type as my main output messages. I once again
>> broadcast-as-in-partitioning the side output stream, unioned it with my
>> processed output from the CoProcessFunction and passed it to my sink,
>> making sure to handle control messages before attempting to do any
>> bucketing.
>>
>>
>>
>> In poor ASCII art, it looks something like the below:
>>
>>
>>
>>
>>
>> _______________ ____________
>>
>> | Schema Source | | Event Source |
>>
>> ----------------------- -------------------
>>
>> | |
>>
>> Broadcast |
>>
>> | __________ |
>>
>> ----- | Processor | -----------
>>
>> | | ----------- Control
>> message side output
>>
>> --------------- |
>>
>> | |
>>
>> | Broadcast
>>
>> | |
>>
>> Union --------------
>>
>> |
>>
>> _______
>>
>> | Sink |
>>
>> -----------
>>
>>
>>
>> I hope this is helpful to someone.
>>
>>
>>
>> Julian
>>
>>
>>
>> *From: *Piotr Nowojski <pn...@apache.org>
>> *Date: *Wednesday, October 14, 2020 at 11:22 PM
>> *To: *"Jaffe, Julian" <Ju...@activision.com>
>> *Cc: *"user@flink.apache.org" <us...@flink.apache.org>
>> *Subject: *Re: Broadcasting control messages to a sink
>>
>>
>>
>> Hi Julian,
>>
>>
>>
>> I think the problem is that BroadcastProcessFunction and SinkFunction
>> will be executed by separate operators, so they won't be able to share
>> state. If you can not split your logic into two, I think you will have to
>> workaround this problem differently.
>>
>>
>>
>> 1. Relay on operator chaining and wire both of them together.
>>
>>
>>
>> If you set up your BroadcastProcessFunction and SinkFunction one after
>> another, with the same parallelism, with the default chaining, without any
>> rebalance/keyBy in between, you can be sure they will be chained together.
>> So the output type of your record between BroadcastProcessFunction and
>> SinkFunction, can be a Union type, of a) your actual payload, b)
>> broadcasted message. Upon initialization/before processing first record, if
>> you have any broadcast state, you would need to forward it's content to the
>> downstream SinkFunction as well.
>>
>>
>>
>> 2. Another solution is that maybe you can try to embed SinkFunction
>> inside the BroadcastProcessFunction? This will require some
>> careful proxying and wrapping calls.
>>
>> 3. As always, you can also write a custom operator that will be doing the
>> same thing.
>>
>>
>>
>> For the 2. and 3. I'm not entirely sure if there are some gotchas that I
>> haven't thought through (state handling?), so if you can make 1. work for
>> you, it will probably be a safer route.
>>
>>
>>
>> Best,
>>
>> Piotrek
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> śr., 14 paź 2020 o 19:42 Jaffe, Julian <Ju...@activision.com>
>> napisał(a):
>>
>> Thanks for the suggestion Piotr!
>>
>>
>>
>> The problem is that the sink needs to have access to the schema (so that
>> it can write the schema only once per file instead of record) and thus
>> needs to know when the schema has been updated. In this proposed
>> architecture, I think the sink would still need to check each record to see
>> if the current schema matches the new record or not? The main problem I
>> encountered when playing around with broadcast state was that I couldn’t
>> figure out how to access the broadcast state within the sink, but perhaps I
>> just haven’t thought about it the right way. I’ll meditate on the docs
>> further 🙂
>>
>>
>>
>> Julian
>>
>>
>>
>> *From: *Piotr Nowojski <pn...@apache.org>
>> *Date: *Wednesday, October 14, 2020 at 6:35 AM
>> *To: *"Jaffe, Julian" <Ju...@activision.com>
>> *Cc: *"user@flink.apache.org" <us...@flink.apache.org>
>> *Subject: *Re: Broadcasting control messages to a sink
>>
>>
>>
>> Hi Julian,
>>
>>
>>
>> Have you seen Broadcast State [1]? I have never used it personally, but
>> it sounds like something you want. Maybe your job should look like:
>>
>>
>>
>> 1. read raw messages from Kafka, without using the schema
>>
>> 2. read schema changes and broadcast them to 3. and 5.
>>
>> 3. deserialize kafka records in BroadcastProcessFunction by using
>> combined 1. and 2.
>>
>> 4. do your logic o
>>
>> 5. serialize records using schema in another BroadcastProcessFunction by
>> using combined 4. and 2.
>>
>> 6. write raw records using BucketingSink
>>
>> ?
>>
>>
>>
>> Best,
>>
>> Piotrek
>>
>>
>>
>> [1]
>> https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html
>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__ci.apache.org_projects_flink_flink-2Ddocs-2Dstable_dev_stream_state_broadcast-5Fstate.html&d=DwMFaQ&c=qE8EibqjfXM-zBfebVhd4gtjNZbrDcrKYXvb1gt38s4&r=zKznthi6OTKpoJID9dIcyiJ28NX59JIQ2bD246nnMac&m=0fL33mv_n-SUiL8AARIrGXmY1d8pdhu4ivDeRjg5f84&s=RjsXnxEVCBz2BGLxe89FU_SpbtfTlRkjsT5J-gbvqFI&e=>
>>
>>
>>
>> śr., 14 paź 2020 o 11:01 Jaffe, Julian <Ju...@activision.com>
>> napisał(a):
>>
>> Hey all,
>>
>>
>>
>> I’m building a Flink app that pulls in messages from a Kafka topic and
>> writes them out to disk using a custom bucketed sink. Each message needs to
>> be parsed using a schema that is also needed when writing in the sink. This
>> schema is read from a remote file on a distributed file system (it could
>> also be fetched from a service). The schema will be updated very
>> infrequently.
>>
>>
>>
>> In order to support schema evolution, I have created a custom source that
>> occasionally polls for updates and if it finds one parses the new schema
>> and sends a message containing the serialized schema. I’ve connected these
>> two streams and then use a RichCoFlatMapFunction to flatten them back into
>> a single output stream (schema events get used to update the parser,
>> messages get parsed using the parser and emitted).
>>
>>
>>
>> However, I need some way to communicate the updated schema to every task
>> of the sink. Simply emitting a control message that is ignored when writing
>> to disk means that only one sink partition will receive the message and
>> thus update the schema. I thought about sending the control message as side
>> output and then broadcasting the resulting stream to the sink alongside the
>> processed event input but I couldn’t figure out a way to do so. For now,
>> I’m bundling the schema used to parse each event with the event, storing
>> the schema in the sink, and then checking every event’s schema against the
>> stored schema but this is fairly inefficient. Also, I’d like to eventually
>> increase the types of control messages I can send to the sink, some of
>> which may not be idempotent. Is there a better way to handle this pattern?
>>
>>
>> (Bonus question: ideally, I’d like to be able to perform an action when
>> all sink partitions have picked up the new schema. I’m not aware of any way
>> to emit metadata of this sort from Flink tasks beyond abusing the metrics
>> system. This approach still leaves open the possibility of tasks picking up
>> the new schema and then crashing for unrelated reasons thus inflating the
>> count of tasks using a specific schema and moreover requires tracking at
>> least the current level of parallelism and probably also Flink task state
>> outside of Flink. Are there any patterns for reporting metadata like this
>> to the job manager?)
>>
>>
>>
>> I’m using Flink 1.8.
>>
>>
--
Thanks & Regards,
Anuj Jain
Mob. : +91- 8588817877
Skype : anuj.jain07
<http://www.oracle.com/>
<http://www.cse.iitm.ac.in/%7Eanujjain/>
Re: Broadcasting control messages to a sink
Posted by "Jaffe, Julian" <Ju...@activision.com>.
Hey AJ,
Depending on your control messages and what you’re trying to accomplish you can simplify the application even further by stripping out the second broadcast and letting operator chaining guarantee that control messages flow appropriately. This results in
_______________ ____________
| Schema Source | | Event Source |
----------------------- -------------------
| |
Broadcast |
| __________ |
----- | Processor | -----------
| |
---------------
|
_______
| Sink |
-----------
Whatever your final topology ends up being, your sink needs to know about the updated schema. If you aren’t using a schema registry, this means you need some way to distinguish between events and control messages in the invoke method of your sink, for example:
@Override
public void invoke(ValueType value, Context ctx) throws Exception {
if (value.isControlMessage()) {
updateSchema(value);
return;
}
writeValue(value)
}
I don’t know what types you’re using etc. and how you’re storing output, so I can’t be much more concrete than this, but if it helps, here’s a simplified version of what I’m doing (I’m reading in avro events and writing them out in bucketed files):
@Override
public void invoke(GenericRecord value, Context ctx) throws Exception {
if (value != null && value.getSchema().equals(CONTROL_SCHEMA)) {
/*
* My schema update control messages don’t actually contain a new schema. This function
* fetches the schema from a remote location, performs various checks and updates various
* systems, and then sets the working schema to the new schema.
*/
updateSchema(value);
return;
}
// … determining the correct bucket for the provided value and initializing the bucketer if necessary …
if (!bucketState.schema.equals(this.schema)) {
bucketState.refreshSchema(this.schema, this.schemaTimestamp);
openNewPartFile(bucketPath, bucketState);
} else if (shouldRoll(bucketState, currentProcessingTime)) {
openNewPartFile(bucketPath, bucketState);
}
bucketState.fileWriter.append(value);
}
I’ve tried to use the method names that match the BucketingSink interface<https://github.com/apache/flink/blob/release-1.11/flink-connectors/flink-connector-filesystem/src/main/java/org/apache/flink/streaming/connectors/fs/bucketing/BucketingSink.java> as much as possible to aid comprehension. I’ve also stripped out a lot of internal logic, instrumentation, sanity checks, etc. The main idea is that
1. Your sink needs to keep track of your schema and update it in response to control messages (in the simplest case, the control messages can just contain the new schema)
2. You need some way to refresh the schema used by your actual writers, which may involve closing current outputs and creating new ones
3. Your schema needs to be serializable so that it can be snapshotted in checkpoints and restored, or you need some way to determine the correct schema in the open() or initializeState() methods
I hope this helps!
For anyone reading this who might be using Avro GenericRecords, note that Flink doesn’t serialize GenericRecords efficiently. You’ll want to write a custom serializer on the pattern of http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/Serialization-performance-td12019.html, the eBay GenericAvroSerializer for Spark, or the many other GenericRecord serializers out there that replace the full schema with the schema’s fingerprint. If your schema is very large, you’ll want to go further and work on schema hashcodes instead of fingerprints (with appropriate safeguards!) because fingerprinting and checking RecordSchemas for equality quickly becomes expensive. In the sample code above, I include the check `(!bucketState.schema.equals(this.schema))`. Make sure that you’re actually comparing schema fingerprints or the like instead of directly calling schema.equals(otherSchema).
Julian
From: aj <aj...@gmail.com>
Date: Friday, December 4, 2020 at 7:20 AM
To: Piotr Nowojski <pn...@apache.org>
Cc: "Jaffe, Julian" <Ju...@activision.com>, "user@flink.apache.org" <us...@flink.apache.org>
Subject: Re: Broadcasting control messages to a sink
Hi Jafee,
Can u please help me out with the sample code how you have written the custom sink and how you using this broadcast pattern to update schema at run time. It will help me.
On Sat, Oct 17, 2020 at 1:55 PM Piotr Nowojski <pn...@apache.org>> wrote:
Hi Julian,
Glad to hear it worked! And thanks for coming back to us :)
Best,
Piotrek
sob., 17 paź 2020 o 04:22 Jaffe, Julian <Ju...@activision.com>> napisał(a):
Hey Piotr,
Thanks for your help! The main thing I was missing was the .broadcast partition operation on a stream (searching for “broadcasting” obviously brought up the broadcast state pattern). This coupled with my misunderstanding of an error in my code as being an error in Flink code resulted in me making this a much harder problem than it needed to be.
For anyone who may find this in the future, Piotr’s suggestion is pretty spot-on. I wound up broadcasting (as in the partitioning strategy) my schema stream and connecting it to my event stream. I then processed those using a CoProcessFunction, using the schema messages to update the parsing for the events. I also emitted a side output message when I processed a new schema, using the same type as my main output messages. I once again broadcast-as-in-partitioning the side output stream, unioned it with my processed output from the CoProcessFunction and passed it to my sink, making sure to handle control messages before attempting to do any bucketing.
In poor ASCII art, it looks something like the below:
_______________ ____________
| Schema Source | | Event Source |
----------------------- -------------------
| |
Broadcast |
| __________ |
----- | Processor | -----------
| | ----------- • Control message side output
--------------- |
| |
| Broadcast
| |
Union --------------
|
_______
| Sink |
-----------
I hope this is helpful to someone.
Julian
From: Piotr Nowojski <pn...@apache.org>>
Date: Wednesday, October 14, 2020 at 11:22 PM
To: "Jaffe, Julian" <Ju...@activision.com>>
Cc: "user@flink.apache.org<ma...@flink.apache.org>" <us...@flink.apache.org>>
Subject: Re: Broadcasting control messages to a sink
Hi Julian,
I think the problem is that BroadcastProcessFunction and SinkFunction will be executed by separate operators, so they won't be able to share state. If you can not split your logic into two, I think you will have to workaround this problem differently.
1. Relay on operator chaining and wire both of them together.
If you set up your BroadcastProcessFunction and SinkFunction one after another, with the same parallelism, with the default chaining, without any rebalance/keyBy in between, you can be sure they will be chained together. So the output type of your record between BroadcastProcessFunction and SinkFunction, can be a Union type, of a) your actual payload, b) broadcasted message. Upon initialization/before processing first record, if you have any broadcast state, you would need to forward it's content to the downstream SinkFunction as well.
2. Another solution is that maybe you can try to embed SinkFunction inside the BroadcastProcessFunction? This will require some careful proxying and wrapping calls.
3. As always, you can also write a custom operator that will be doing the same thing.
For the 2. and 3. I'm not entirely sure if there are some gotchas that I haven't thought through (state handling?), so if you can make 1. work for you, it will probably be a safer route.
Best,
Piotrek
śr., 14 paź 2020 o 19:42 Jaffe, Julian <Ju...@activision.com>> napisał(a):
Thanks for the suggestion Piotr!
The problem is that the sink needs to have access to the schema (so that it can write the schema only once per file instead of record) and thus needs to know when the schema has been updated. In this proposed architecture, I think the sink would still need to check each record to see if the current schema matches the new record or not? The main problem I encountered when playing around with broadcast state was that I couldn’t figure out how to access the broadcast state within the sink, but perhaps I just haven’t thought about it the right way. I’ll meditate on the docs further 🙂
Julian
From: Piotr Nowojski <pn...@apache.org>>
Date: Wednesday, October 14, 2020 at 6:35 AM
To: "Jaffe, Julian" <Ju...@activision.com>>
Cc: "user@flink.apache.org<ma...@flink.apache.org>" <us...@flink.apache.org>>
Subject: Re: Broadcasting control messages to a sink
Hi Julian,
Have you seen Broadcast State [1]? I have never used it personally, but it sounds like something you want. Maybe your job should look like:
1. read raw messages from Kafka, without using the schema
2. read schema changes and broadcast them to 3. and 5.
3. deserialize kafka records in BroadcastProcessFunction by using combined 1. and 2.
4. do your logic o
5. serialize records using schema in another BroadcastProcessFunction by using combined 4. and 2.
6. write raw records using BucketingSink
?
Best,
Piotrek
[1] https://ci.apache.org/projects/flink/flink-docs-stable/dev/stream/state/broadcast_state.html<https://urldefense.proofpoint.com/v2/url?u=https-3A__ci.apache.org_projects_flink_flink-2Ddocs-2Dstable_dev_stream_state_broadcast-5Fstate.html&d=DwMFaQ&c=qE8EibqjfXM-zBfebVhd4gtjNZbrDcrKYXvb1gt38s4&r=zKznthi6OTKpoJID9dIcyiJ28NX59JIQ2bD246nnMac&m=0fL33mv_n-SUiL8AARIrGXmY1d8pdhu4ivDeRjg5f84&s=RjsXnxEVCBz2BGLxe89FU_SpbtfTlRkjsT5J-gbvqFI&e=>
śr., 14 paź 2020 o 11:01 Jaffe, Julian <Ju...@activision.com>> napisał(a):
Hey all,
I’m building a Flink app that pulls in messages from a Kafka topic and writes them out to disk using a custom bucketed sink. Each message needs to be parsed using a schema that is also needed when writing in the sink. This schema is read from a remote file on a distributed file system (it could also be fetched from a service). The schema will be updated very infrequently.
In order to support schema evolution, I have created a custom source that occasionally polls for updates and if it finds one parses the new schema and sends a message containing the serialized schema. I’ve connected these two streams and then use a RichCoFlatMapFunction to flatten them back into a single output stream (schema events get used to update the parser, messages get parsed using the parser and emitted).
However, I need some way to communicate the updated schema to every task of the sink. Simply emitting a control message that is ignored when writing to disk means that only one sink partition will receive the message and thus update the schema. I thought about sending the control message as side output and then broadcasting the resulting stream to the sink alongside the processed event input but I couldn’t figure out a way to do so. For now, I’m bundling the schema used to parse each event with the event, storing the schema in the sink, and then checking every event’s schema against the stored schema but this is fairly inefficient. Also, I’d like to eventually increase the types of control messages I can send to the sink, some of which may not be idempotent. Is there a better way to handle this pattern?
(Bonus question: ideally, I’d like to be able to perform an action when all sink partitions have picked up the new schema. I’m not aware of any way to emit metadata of this sort from Flink tasks beyond abusing the metrics system. This approach still leaves open the possibility of tasks picking up the new schema and then crashing for unrelated reasons thus inflating the count of tasks using a specific schema and moreover requires tracking at least the current level of parallelism and probably also Flink task state outside of Flink. Are there any patterns for reporting metadata like this to the job manager?)
I’m using Flink 1.8.
--
Thanks & Regards,
Anuj Jain
Mob. : +91- 8588817877
Skype : anuj.jain07