FLINK-6117 issue work around

Hi, Using flink 1.2.0, I faced to issue https://issues.apache.org/jira/browse/FLINK-6117 https://issues.apache.org/jira/browse/FLINK-6117. This issue is fixed at version 1.3.0. But I have some reason to trying to find out work around. I did, 1. change source according to

Re: Fwd: HA : My job didn't restart even if task manager restarted.

I am still struggling to solve this problem. I have no doubt that the JOB should automatically restart after restarting the TASK MANAGER in YARN MODE. Is it a misunderstood? Problem seems that *JOB MANAGER still try to connect to old TASK MANAGER even after new TASK MANAGER container be created.*

Flink on AWS EMR Protobuf

Hello, Has anybody experienced the following error on AWS EMR 5.8.0 with Flink 1.3.1 java.lang.ClassCastException: org.apache.hadoop.hdfs.protocol.proto.ClientNamenodeProtocolProtos$GetFileInfoRequestProto cannot be cast to com.google.protobuf.Message Thanks,

Broadcast Config through Connected Stream

Hi All, I looked into an earlier email about the topic broadcast config through connected stream and I couldn't find the conclusion. I can't do the below approach since I need the config to be published to all operator instances but I need keyed state for external querying.

RE: DataSet: partitionByHash without materializing/spilling the entire partition?

We have the same issue. We are finding that we cannot express the data flow in a natural way because of unnecessary spilling. Instead, we're making our own operators which combine multiple steps together and essentially hide it from flink OR sometimes we even have to read an input dataset once

Re: State Maintenance

Thanks Gordon for your response. I have around 80 parallel flatmap operator instances and each instance requires 3 states. Out of which one is user state in which each operator will have unique user's data and I need this data to be queryable. The other two states are kind of static states which

Re: Process Function

Thanks a lot everyone. I have the user data ingested from kafka and it is keyed by userid. There are around 80 parallel flatmap operator instances after keyby and there are around few million users. The map state includes userid as the key and some value. I guess I will try the approach that

Re: Does RocksDB need a dedicated CPU?

Thank you, Kien! On Tue, Sep 5, 2017 at 8:01 AM, Kien Truong wrote: > Hi, > > In my experience, RocksDB uses very little CPU, and doesn't need a > dedicated CPU. > > However, it's quite disk intensive. You'd need fast, ideally dedicated > SSDs to achieve the best

Re: Process Function

Hi, This is mostly correct, but you cannot register a timer in open() because we don't have an active key there. Only in process() and onTimer() can you register a timer. In your case, I would suggest to somehow clamp the timestamp to the nearest 2 minute (or whatever) interval or to keep an

Re: DataSet: CombineHint heuristics

Hi Gábor, thank you very much for your explanation, that makes a lot of sense. Best regards, Urs On 05.09.2017 14:32, Gábor Gévay wrote: > Hi Urs, > > Yes, the 1/10th ratio is just a very loose rule of thumb. I would > suggest to try both the SORT and HASH strategies with a workload that > is

Re: Does RocksDB need a dedicated CPU?

Hi, In my experience, RocksDB uses very little CPU, and doesn't need a dedicated CPU. However, it's quite disk intensive. You'd need fast, ideally dedicated SSDs to achieve the best performance. Regards, Kien On 9/5/2017 1:15 PM, Bowen Li wrote: Hi guys, Does RocksDB need a dedicated

Re: Process Function

Hi Navneeth, Currently, I don't think there is any built-in functionality to trigger onTimer periodically. As for the second part of your question, do you mean that you want to query on which key the fired timer was registered from? I think this also isn't possible right now. I'm looping in

Shuffling between map and keyBy operator

I have a streaming application that has a keyBy operator followed by an operator working on the keyed values (a custom sum operator). If the map operator and aggregate operator are running on same Task Manager , will Flink always serialize and deserialize the tuples, or is there an optimization

Task Manager was lost/killed due to full GC

Hi, I'm running a Flink batch job that reads almost 1 TB of data from S3 and then performs operations on it. A list of filenames are distributed among the TM's and each subset of files is read from S3 from each TM. This job errors out at the read step due to the following error:

Re: LatencyMarker

Hi! Yes, backpressure should also increase the latency value calculated from LatencyMarkers. LatencyMarkers are special events that flow along with the actual stream records, so they should also be affected by backpressure. Are you asking because you observed otherwise? Cheers, Gordon --

Re: State Maintenance

Hi Navneeth, Answering your three questions separately: 1. Yes. Your MapState will be backed by RocksDB, so when removing an entry from the map state, the state will be removed from the local RocksDB as well. 2. If state classes are not POJOs, they will be serialized by Kryo, unless a custom

Re: DataSet: CombineHint heuristics

Hi Urs, Yes, the 1/10th ratio is just a very loose rule of thumb. I would suggest to try both the SORT and HASH strategies with a workload that is as similar as possible to your production workload (similar data, similar parallelism, etc.), and see which one is faster for your specific use case.

Re: How to flush all window states after Kafka (0.10.x) topic was removed

Hi Tony, Currently, the functionality that you described does not exist in the consumer. When a topic is deleted, as far as I know, the consumer would simply consider the partitions as unreachable and continue to try fetching records from them until they are up again. I'm not entirely sure if a

Re: termination of stream#iterate on finite streams

Hello, There is a Flink Improvement Proposal to redesign the iterations: https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=66853132 This will address the termination issue. Best, Gábor On Mon, Sep 4, 2017 at 11:00 AM, Xingcan Cui wrote: > Hi Peter, > >

DataSet: partitionByHash without materializing/spilling the entire partition?

Hi all, we have a DataSet pipeline which reads CSV input data and then essentially does a combinable GroupReduce via first(n). In our first iteration (readCsvFile -> groupBy(0) -> sortGroup(0) -> first(n)), we got a jobgraph like this: source --[Forward]--> combine --[Hash Partition on 0,

Re: Process Function

How are you determining your data is stale? Also if you want to know the key, why don't you store the key in your state as well? -- Sent from: http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/

How to flush all window states after Kafka (0.10.x) topic was removed

Hi, I have a simple streaming job consuming data from Kafka and use time window to aggregate them. I am wondering if there is a built-in function to send a max watermark when consumer find this topic is not available, so that the window function can flush all state to the sink function. My Kafka

Does RocksDB need a dedicated CPU?

Hi guys, Does RocksDB need a dedicated CPU? Do we need to allocate one CPU for each RocksDB while deploying Flink cluster with RocksDB state backend? I think there's probably no need since RocksDB is a native 'library', but I want to confirm it with Flink community. Thanks, Bowen