You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@pulsar.apache.org by Apache Pulsar Slack <ap...@gmail.com> on 2018/05/10 09:11:02 UTC
Slack digest for #general - 2018-05-10
2018-05-09 09:32:20 UTC - Byron: @Idan I setup a cluster on k8s, no tuning (frankly I am completely new to tuning things on the JVM) and easily achieve your throughput requirements.
----
2018-05-09 09:32:34 UTC - Byron: @Byron uploaded a file: <https://apache-pulsar.slack.com/files/UACD54WB1/FALNQVA8J/pulsar-dash.png|pulsar-dash.png>
----
2018-05-09 09:33:13 UTC - Byron: the irregular patterns are because i am still testing workloads on it, but it has been running for several days without issue so far.
----
2018-05-09 09:34:48 UTC - Byron: just mentioning my anecdote to hopefully reduce your doubt
+1 : Sijie Guo, Ali Ahmed, Matteo Merli
----
2018-05-09 09:53:56 UTC - Idan: Thank you guys. Ill take this under considerations
----
2018-05-09 13:34:20 UTC - William Fry: @William Fry has joined the channel
----
2018-05-09 15:33:24 UTC - Rob V: Hi there, regarding Pulsar Functions, the only comparable solution that I know of is Kafka Streams, where you deploy a Java application that does the message transformation. In my experience running Kafka and Kafka Streams in production under significant workload is that you do have to be careful configuring your application (memory/CPU/Kafka consumer/producer fine-tuning). Although Functions sound like an interesting idea, how does it supposed to scale and make sure that functions running on the cluster won't have an impact on the cluster itself? Also, does it support state handling like Kafka Streams does?
----
2018-05-09 15:41:38 UTC - George Goranov: @George Goranov has joined the channel
----
2018-05-09 16:39:01 UTC - Matteo Merli: > Although Functions sound like an interesting idea, how does it supposed to scale and make sure that functions running on the cluster won’t have an impact on the cluster itself?
@Rob V There are different deployment options :
* Local runner (similar to Kafka streams, you start your instances independently)
* Managed mode — There’s a pool of workers that executes instances of the functions. The workers can run as part of broker (make sense for a small single-tenant cluster) or as a separate service layer. Inside each worker, function instances can run as threads, processes or, in near future, as containers
----
2018-05-09 16:44:07 UTC - Rob V: thank you @Matteo Merli
----
2018-05-09 17:13:32 UTC - Sebastian Schepens: @Sebastian Schepens has joined the channel
----
2018-05-09 17:44:08 UTC - Karthik Palanivelu: Hi there, I am getting the following error when starting standalone in RHEL7 box. It is snot using the IP from Client.conf and using 127.0.0.1 which is failing in this case. Can you please let me know which setting need to be changed for that IP?
----
2018-05-09 17:44:15 UTC - Karthik Palanivelu: @Karthik Palanivelu uploaded a file: <https://apache-pulsar.slack.com/files/U7VRE0Q1G/FALFZ3Z33/-.lua|Untitled>
----
2018-05-09 17:51:15 UTC - Karthik Palanivelu: @Karthik Palanivelu uploaded a file: <https://apache-pulsar.slack.com/files/U7VRE0Q1G/FAN45KQ23/-.php|Untitled>
----
2018-05-09 17:53:17 UTC - Matteo Merli: @Karthikeyan Palanivelu In this case it is the functions worker embedded in the standalone service. You can disable functions worker, if you’re not using that, by passing the `--no-functions-worker` to the `pulsar standalone` command
----
2018-05-09 18:05:54 UTC - Karthik Palanivelu: @Matteo Merli That works, Is there a separate config to add the IP or it should read from client.conf? Please let me know where it is picking the 127.0.0.1 from?
----
2018-05-09 18:24:13 UTC - Matteo Merli: The function worker is trying to use some topics for its internal management. In standalone mode we assume connecting to 127.0.0.1 is always working (which might not be true in your case)
----
2018-05-09 18:34:24 UTC - Daniel Ferreira Jorge: Hi, I'm running a small pulsar cluster in production on kubernetes (GKE). It has 5 bookies (one bookie per node), each one with a 3TB Persistent disk. The brokers are in another set of nodes, so is the zookeeper cluster. The write quorum is 3 and ack quorum is 2. I need to migrate this 5 bookies to a new set of 5 nodes, without downtime and without impacting producers and consumers. I'm trying to avoid having to use auto-recovery since this will have to copy the data of each old bookie to each new bookie and this will take a long time. This procedure has to be done periodically (roughly every 3 months) so the worker nodes in GKE can keep up with newer kubernetes versions on the master nodes.
If I just delete a bookie pod so it will be rescheduled on another node by kubernetes (a procedure that takes roughly a minute due to the migration of the persistent disk to the new node) will this be enough? Will it be safe? Will the bookie come back online and rejoin the cluster "gracefully" since it will keep the data?
Do you guys recommend a better way to do a "rolling migration" of bookie pods to new kubernetes nodes?
----
2018-05-09 18:45:54 UTC - Matteo Merli: > If I just delete a bookie pod so it will be rescheduled on another node by kubernetes (a procedure that takes roughly a minute due to the migration of the persistent disk to the new node) will this be enough? Will it be safe? Will the bookie come back online and rejoin the cluster “gracefully” since it will keep the data?
> Do you guys recommend a better way to do a “rolling migration” of bookie pods to new kubernetes nodes?
Since the data is stored in the persistent volume, the bookies pods become essentially stateless, so you’re right that auto-recovery is not needed.
The only thing that needs to be ensured is that the new bookie Pod identifies itself with the same “name” as the old bookie. That is required so that BK clients can connect to the right Pod and assume it still has the data.
Bookies have a “cookie” verification mechanism that will prevent a bookie to impersonate a different one. One such example would be a new “bookie-2” trying to start with the same persistent volume that was previously used by “bookie-1".
To ensure the same identifier is used, in K8S the easiest way is to use StatefulSet for bookies and configure them to register and advertise with the “hostname” rather than the “IP address” (which is the default). Since the Pod IP can change when Pod is rescheduled to a different node, but the Pod name will remain the same “bookie-1”. That should already be configured that way in the example K8S deployment specs: <https://github.com/apache/incubator-pulsar/blob/master/deployment/kubernetes/google-kubernetes-engine/bookie.yaml#L53>
----
2018-05-09 18:51:09 UTC - Daniel Ferreira Jorge: I use a statefulset for the bookies. So its safe to simply delete the pod and wait for it to be rescheduled, great! One question regarding that: When we delete a pod on kubernetes, it will send a SIGTERM to the pod to terminate the main process and wait for the grace period and then send a SIGKILL. Does the bookeeper handle the SIGTERM to stop itself gracefully?
----
2018-05-09 18:52:52 UTC - Daniel Ferreira Jorge: Also, after I delete my first bookie and it come back online, how can I verify that everything is OK? (if it rejoined the cluster, if there are no under replicated ledgers, etc)
----
2018-05-09 18:53:00 UTC - Matteo Merli: I believe it should do the grace shutdown on sigterm, flush all write cache and then exist
----
2018-05-09 18:54:29 UTC - Ali Ahmed: kubernetes passes sigterm to the pid 1, I don’t think bookie is running with that
----
2018-05-09 18:55:51 UTC - Ali Ahmed: here are some docs here
<https://pracucci.com/graceful-shutdown-of-kubernetes-pods.html>
----
2018-05-09 18:59:59 UTC - Matteo Merli: > Also, after I delete my first bookie and it come back online, how can I verify that everything is OK? (if it rejoined the cluster, if there are no under replicated ledgers, etc)
Good question. The auto-recovery itself doesn’t verify all the data, it’s just based on the meta-data though it double-checks few entries per each ledger.
I think you can indeed use auto-recovery, perhaps manually to force the double-check
----
2018-05-09 19:00:09 UTC - Joe Francis: @Daniel Ferreira Jorge @Matteo Merli Wouldn't making the existing 5 bookies read-only and adding .5 new bookies be simpler?
----
2018-05-09 19:00:33 UTC - Joe Francis: The read-only bookies can be dropped once the ledgers roll-off
----
2018-05-09 19:01:06 UTC - Daniel Ferreira Jorge: @Ali Ahmed Hmm... I will have to check that! Thanks
----
2018-05-09 19:01:38 UTC - Matteo Merli: Like `bin/bookkeeper shell recover $SRC_BOOKIE`
----
2018-05-09 19:02:55 UTC - Matteo Merli: In any case, for the restart. If it’s not graceful shutdown, it just means it will replay some entries from the journal
----
2018-05-09 19:04:45 UTC - Daniel Ferreira Jorge: @Matteo Merli Ok that is great! So to check each bookie I can do a recovery just to be sure, correct?
----
2018-05-09 19:06:13 UTC - Daniel Ferreira Jorge: @Joe Francis I could not understand how would this work... could you explain it please?
----
2018-05-09 19:08:28 UTC - Joe Francis: Add 5 new bookies. Then making the existing ones read only. New data will get written to new bookies, since the old ones are now read only. Once all consumers are done with reading data on old ledgers, those ledgers will get deleted. At that point you can retire the old bookies.
----
2018-05-09 19:09:18 UTC - Daniel Ferreira Jorge: That would not work for me, because I keep the data for re-consumption forever
----
2018-05-09 19:10:06 UTC - Matteo Merli: @Joe Francis yes, that would be the best strategy for on-prem bare metal. but if you’re in cloud with persistent volumes it’s easier to just restart the Pod in a different node
----
2018-05-09 19:11:39 UTC - Daniel Ferreira Jorge: I need only to change the underlying node that the pod is running but when the pod come back online, everything should be the same as it was before.
----
2018-05-09 19:12:09 UTC - Daniel Ferreira Jorge: @Matteo Merli @Joe Francis @Ali Ahmed Thank you very much for all the help
----
2018-05-09 19:27:54 UTC - Byron: :point_up: great information
----
2018-05-10 01:55:36 UTC - Senthilkumar Kalaiselvan: <!here> , Hi All , Good Morning... Looking into <https://pulsar.incubator.apache.org/> this doc , Seems to be Apache Pulsar is a good candidate for my streaming use case ... Can someone help me to understand how it differs from Kafka ? Any documentation with benchmarking numbers ?
----
2018-05-10 02:03:54 UTC - Ali Ahmed: In simple terms pulsar could be considered a 2nd generation distributed pub sub, it provides a superset of kafka capabilities and has a kafka compatibility api, if you can describe your use case in more detail we can provide more guidance.
----
2018-05-10 02:09:33 UTC - Senthilkumar Kalaiselvan: At high level , We collect different data from servers ( timestamp is the key ) and push it to Message Buffer(Kafka or Pulsar ?). While consuming back the same data should be consumed based on timestamp i.e strong ordering ... also sometime we want to move consumer position to some timestamp .. for example : start consuming from now -3 hours , now - 1 hour etc ...
Traffic : 250K Requests/ Sec each request : ~1.2 KB .
----
2018-05-10 02:09:59 UTC - Senthilkumar Kalaiselvan: Looking to see the benchmarking numbers if any ...
----
2018-05-10 02:10:44 UTC - Ali Ahmed: kafka and pulsar have already been benchmarked.
<http://openmessaging.cloud/docs/benchmarks/>
----
2018-05-10 02:11:28 UTC - Ali Ahmed: pulsar today is bout 4x faster , with 2.0 on the horizon it will get a lot faster
----
2018-05-10 02:11:58 UTC - Ali Ahmed: pulsar has much stronger data durability and ordering guarantees then Kafka
----
2018-05-10 02:12:32 UTC - Senthilkumar Kalaiselvan: Awesome! I'll get started on building POC..
----
2018-05-10 02:12:58 UTC - Ali Ahmed: your request volume is modest so it’s not a problem at all for a mid size clsuter
----
2018-05-10 02:13:08 UTC - Ali Ahmed: no specific tuning is necessary
----
2018-05-10 02:13:49 UTC - Senthilkumar Kalaiselvan: 250MBps is starting load , it can grow up 500MB in 2019 Q1 .
----
2018-05-10 02:15:21 UTC - Ali Ahmed: that’s not as problem , you just add nodes if needed , there is no rebalancing or downtime in pulsar, it just scales horzontally
----
2018-05-10 02:15:55 UTC - Senthilkumar Kalaiselvan: nice .
----
2018-05-10 02:18:50 UTC - Ali Ahmed: I recommend doing the POC with 2.0 RC , it will go GA within within a few weeks.
----
2018-05-10 02:19:09 UTC - Senthilkumar Kalaiselvan: Cool, Thanks.
----