storm-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Stig Rohde Døssing <s...@apache.org>
Subject Re: Migrating from storm-kafka to storm-kafka-client
Date Fri, 12 Jan 2018 19:22:37 GMT
There's a small discrepancy, in the old spout we track the offset the next
batch should start at, while in storm-kafka-client we track the last
emitted offset. So for example if offset 0, 1, 2 were emitted, the old
spout would store offset = 0, nextOffset = 3. The new spout would store
firstOffset = 0, lastOffset = 2. I think it should be lastOffset =
nextOffset - 1. Other than that I agree with your mapping.

2018-01-12 3:43 GMT+01:00 Nasron Cheong <nasron@gmail.com>:

> Hi Stig,
>
> That's great! Thanks for all the info. Looking through the code, one small
> detail is the difference between storm-kafka-client's format and
> storm-kafka. The former uses 'firstOffset' and 'lastOffset' and the latter
> uses 'offset' and 'nextOffset'.
>
> So, can I map with
>
> firstOffset = offset
>
> and
>
> lastOffset = nextOffset+1 ?
>
> Looking through the code it seems to be that nextOffset is placed after
> the last consumed message, but I'm not sure.
>
> Thanks
>
> - Nasron
>
> On Thu, Jan 11, 2018 at 5:43 PM, Stig Rohde Døssing <srdo@apache.org>
> wrote:
>
>> Nasron,
>>
>> Okay, migrating a Trident spout is a very different thing. Trident spouts
>> store their state in Storm's zookeeper (unless you decide otherwise by
>> setting transactional.zookeeper.servers in storm.yaml). This also
>> applies to the storm-kafka-client Trident spout, so we won't need to move
>> offsets into Kafka.
>>
>> The idea of stopping all the producers and starting at LATEST (or
>> UNCOMMITTED_LATEST) is decent, but as you note there's a (small) risk of
>> skipping tuples. In order to get Trident to commit something, you have to
>> deploy the new topology with LATEST and start the producers again, wait
>> until at least one commit happens, and then take the topology back down and
>> redeploy with whatever your first poll strategy normally is. If the worker
>> crashes before the spout manages to commit something, you will skip tuples.
>>
>> If you don't want to do that, here's my notes on storm-kafka ->
>> storm-kafka-client for Trident:
>>
>> The storage formats and zk paths for the two spouts are a little
>> different. Both spouts store their state as JSON maps, but some of the keys
>> are different. I use ${} below to indicate variable substitution.
>>
>> The root path (in the following: zkRoot) for your spouts data is
>> /${transactional.zookeeper.root from storm.yaml}/${txId you set with
>> TopologyBuilder.newStream}/user.
>>
>> For the storm-kafka spout the offsets are stored in one of the following
>> two paths:
>> ${zkRoot}/${topicName}partition_${partition} if you are using wildcard
>> topic subscriptions
>> ${zkRoot}/partition_${partition} otherwise
>>
>> The storage format for storm-kafka is as follows:
>> { "${topicName}partition_${partition}': {"offset": 0, "nextOffset": 2 }
>> } if you are using wildcard topic subscriptions
>> { "${topicName}partition_${partition}': {"offset": 0, "nextOffset": 2 }
>> } otherwise (I left out some irrelevant properties)
>>
>> For storm-kafka-client the zk path is
>> ${zkRoot}/${topicName}@${partition}
>>
>> and the storage format is
>> { "${topicName}@${partition}': {"firstOffset": 0, "nextOffset": 2 } }
>>
>> In order to migrate from storm-kafka to storm-kafka-client, we need to
>> stop the topology and run a script that moves the offsets from the old
>> location/format to the new location/format. There's no way to tell Trident
>> to read from one path/format and write to another, so it has to be done
>> offline. Once the offsets are migrated, the spout can be replaced in the
>> topology and the topology can be redeployed.
>>
>> I might look at writing an application that can do this at some point,
>> but it might take me a while. If you'd like to look at it yourself, here's
>> some pointers where to start:
>> * This is where the offset are written to Zookeeper, assuming you use an
>> opaque spout https://github.com/apache/storm/blob/master/storm-client/src
>> /jvm/org/apache/storm/trident/spout/OpaquePartitionedTrident
>> SpoutExecutor.java#L184. You might want to look at this class for a bit
>> (particularly the emit function), because it's pretty useful for
>> understanding how/where Trident stores metadata for spouts.
>> * The return value of https://github.com/apache/stor
>> m/blob/master/external/storm-kafka/src/jvm/org/apache/storm/
>> kafka/trident/TridentKafkaEmitter.java#L85 defines the format of what's
>> being saved to Zookeeper for storm-kafka. It's being wrapped in a map so
>> the full written value is { "${topicName}partition_${partition}':
>> ${theReturnValue} } (see the storage format note above, it's different if
>> you're not using wildcard subscriptions)
>> * Similarly for storm-kafka-client the return value of
>> https://github.com/apache/storm/blob/master/external/storm-k
>> afka-client/src/main/java/org/apache/storm/kafka/spout/tride
>> nt/KafkaTridentSpoutEmitter.java#L106 defines the format of what that
>> spout saves to Zookeeper (and expects to find).
>> * You should use zkCli (it's in your zookeeper/bin directory) to explore
>> your Zookeeper filesystem. It should be pretty easy to find your offsets in
>> there with that tool.
>>
>> Sorry about the wall of text, this turned out to have a lot of detail to
>> cover.
>>
>> 2018-01-10 21:40 GMT+01:00 Nasron Cheong <nasron@gmail.com>:
>>
>>> Thanks Stig,
>>>
>>> So after some digging, I realized we are really migrating from the kafka
>>> trident emitter in storm-kafka, to the trident emitter in
>>> storm-kafka-client.
>>>
>>> As far as I can see, the offset information is still stored in zk, and
>>> the offset info for storm-kafka is (https://github.com/apache/sto
>>> rm/blob/master/external/storm-kafka/src/jvm/org/apache/storm
>>> /kafka/trident/TridentKafkaEmitter.java#L140)
>>>
>>> However this seems quite different from storm-kafka-client, which uses
>>> https://github.com/apache/storm/blob/master/external/st
>>> orm-kafka-client/src/main/java/org/apache/storm/kafka/spout/
>>> trident/KafkaTridentSpoutBatchMetadata.java#L56
>>>
>>> I'm not sure under which zknode this information is stored - and if the
>>> zknode itself is different between the two implementations.
>>>
>>> Looks like I need a tool to copy the stored values in zk from old
>>> storm-kafka to storm-kafka-client?
>>>
>>> Another option I suppose is to:
>>> - stop topic producers
>>> - run the old code until it drains all topics
>>> - start new code with FirstPollOffsetStrategy.LATEST
>>>
>>> Although this seems risky.
>>>
>>> Thanks!
>>>
>>> - Nasron
>>>
>>>
>>> On Thu, Dec 21, 2017 at 4:23 PM, Stig Rohde Døssing <srdo@apache.org>
>>> wrote:
>>>
>>>> Hi Nasron,
>>>>
>>>> I don't believe there's currently a tool to help you migrate. We did it
>>>> manually by writing a small utility that looked up the commit offsets in
>>>> Storm's Zookeeper, opened a KafkaConsumer with the new consumer group id
>>>> and committed the offsets for the appropriate partitions. We stopped our
>>>> topologies, used this utility and redeployed with the new spout.
>>>>
>>>> Assuming there isn't already a tool for migration floating around
>>>> somewhere, I think we could probably build some migration support into the
>>>> storm-kafka-client spout. If the path to the old offsets in Storm's
>>>> Zookeeper is given, we might be able to extract them and start up the new
>>>> spout from there.
>>>>
>>>> 2017-12-19 21:59 GMT+01:00 Nasron Cheong <nasron@gmail.com>:
>>>>
>>>>> Hi,
>>>>>
>>>>> I'm trying to determine steps for migration to the storm-kafka-client
>>>>> in order to use the new kafka client.
>>>>>
>>>>> It's not quite clear to me how offsets are migrated - is there a
>>>>> specific set of steps to ensure offsets are moved from the ZK based offsets
>>>>> into the kafka based offsets?
>>>>>
>>>>> Or is the original configuration respected, and storm-kafka-client can
>>>>> mostly be a drop in replacement?
>>>>>
>>>>> I want to avoid having spouts reset to the beginning of topics after
>>>>> deployment, due to this change.
>>>>>
>>>>> Thanks.
>>>>>
>>>>> - Nasron
>>>>>
>>>>
>>>>
>>>
>>
>

Mime
View raw message