Hi penghui and haiting,
I try to figure out how the (anti-)affinity works.
> if I understand correctly, it looks like if we have a partitioned topic
with 10
> partitions under a namespace with 16 bundles, if applies the
anti-affinity policy,
> partition-0 map to bundle 0, partition-1 map to bundle 1, and so on.
> Of course, it is not necessary for every partitioned topic to start from
bundle 0,
> we can use the partition-0 hash to determine the start bundle index.
As penghui described, I think this is a mechanism for assigning topics to a
bundle that controls how a topic is mapped to a bundle.
> IMO, this affinity serves the purpose of isolating an abnormal topic to
some spare
> brokers. These brokers host these kind of topics only. Here are some
cases :
And as haiting mentioned here, It's more like an isolation policy that
decides topics can be owned by which broker.
So, I am a little confused about how the (anti-)affinity works now.
Back to this PIP which aims to solve the problem that a small number of
topics in a bundle have a load that exceeds the average.
So, we can
1. get the positions for the topics with we are interested( need a new API)
2. calculate the position to split this bundle(also need a new API)
I think this way is enough for solving the problem.
And the (anti-)affinity way needs more discussion or maybe we can introduce
a new PIP for it.
What do you think?@penghui @haiting
Thanks,
Aloys
Aloys Zhang <aloyszh...@apache.org> 于2022年2月21日周一 14:39写道：
> Hi, penghui
>
> > The new API does not necessarily have to query by topic one by one,
> we have listed all the "topic -> position" of a bundle?
>
> I see. After we got all the positions of the topics we want to split in a
> bundle, it's quite easy for us to decide how to it.
>
> Haiting Jiang <jianghait...@apache.org> 于2022年2月20日周日 12:05写道：
>
>> > Do you have an example for affinity? I don't fully understand how this
>> is
>> > used
>> > in practice.
>>
>> IMO, this affinity serves the purpose of isolating an abnormal topic to
>> some spare
>> brokers. These brokers host these kind of topics only. Here are some
>> cases :
>>
>> 1. A topic may have unexpected short spike traffic flows periodically and
>> causing broker overloads and negative impact on other topics.
>> Until we have more proper solutions, we can always isolate these topics
>> first,
>> and make the service recover time as small as possible.
>>
>> 2. Some users may encounter some bugs in brokers, and we can isolate the
>> topic to
>> exclusive brokers, and use more radical approach to locate the bug, like
>> enable debug
>> level logs or even add some temporary code patch.
>>
>> 3. User may already have configured failure domain and anti-affinity
>> namespace, but with
>> business logic code changes, some topic may need to migrate from one
>> namespace
>> to another. This will take some time for user to change the client side
>> config.
>> In the meanwhile, we can isolate the topic first.
>>
>> Thanks,
>> Haiting
>>
>> On 2022/02/18 15:26:09 PengHui Li wrote:
>> > Hi Haiting,
>> >
>> > > I think this approach have more potential with abnormal topic
>> isolation.
>> > If we can introduce
>> > some kind of bundle isolation strategy, (like broker-bundle affinity and
>> > anti-affinity mechanism), we can easily isolate some unexpected traffic
>> to
>> > some empty brokers.
>> > IMO, this would improve the stability of broker cluster.
>> >
>> > if I understand correctly, it looks like if we have a partitioned topic
>> > with 10
>> > partitions under a namespace with 16 bundles, if applies the
>> anti-affinity
>> > policy,
>> > partition-0 map to bundle 0, partition-1 map to bundle 1, and so on.
>> > Of course, it is not necessary for every partitioned topic to start from
>> > bundle 0,
>> > we can use the partition-0 hash to determine the start bundle index.
>> >
>> > Do you have an example for affinity? I don't fully understand how this
>> is
>> > used
>> > in practice.
>> >
>> > Best,
>> > Penghui
>> >
>> > On Fri, Feb 18, 2022 at 11:16 PM PengHui Li <peng...@apache.org> wrote:
>> >
>> > > Hi Aloys,
>> > >
>> > > > Do you mean that
>> > > 1. First, add a new API, maybe `getHashPositioin`, to get the hash
>> > > position in a bundle
>> > > 2. Then use this position to split the overloaded bundle
>> > > If so, when we split a bundle with multi partitions of a topic, we
>> need to
>> > > call the `getHashPositioin` multi times to get the middle position of
>> all
>> > > these positions.
>> > >
>> > > Yes, this want I mean. In this way, users can control to assign 1
>> topic or
>> > > 3 topics to one bundle. This is more like increasing the transparency
>> of
>> > > the topic in the bundle, you can all the positions of the topics, so
>> how
>> > > planning for bundle splitting becomes more flexible.
>> > >
>> > > The new API does not necessarily have to query by topic one by one,
>> > > we have listed all the "topic -> position" of a bundle?
>> > >
>> > > Thanks,
>> > > Penghui
>> > >
>> > > On Fri, Feb 18, 2022 at 4:51 PM Haiting Jiang <
>> jianghait...@apache.org>
>> > > wrote:
>> > >
>> > >> Hi Aloys,
>> > >> +1 for this great PIP.
>> > >>
>> > >> > The Admin CLI `bin/pulsar-admin namespaces split-bundle -b
>> > >> ${bundle_range}`
>> > >> > will add a new parameter "--topic" or "-t" for `outstanding topic`
>> > >> name.
>> > >>
>> > >> Do we have limitation on this "topic" parameter. Can this be a
>> > >> partitioned topic?
>> > >> If so, will this new algorithm split the bundle into more than 2
>> bundles?
>> > >> like each bundle for
>> > >> one partition.
>> > >>
>> > >> > This algorithm has a disadvantage, it can only deal
>> > >> > with one `outstanding topic`.
>> > >>
>> > >> For this disadvantage, I think it can be solved by extends the
>> "topic"
>> > >> parameter from one topic to a topic list.
>> > >>
>> > >> > The other algorithm is to split the bundle at the hashcode point
>> of the
>> > >> > `outstanding partition` which will split the bundle into three
>> bundles
>> > >> once
>> > >> > a time. The middle one contains the only point the hashcode of the
>> > >> > `outstanding partition, the left one is less than the hashcode, the
>> > >> right
>> > >> > one is more than the hashcode.
>> > >> > E.g. if we have a bundle 0x00_0x10 contains two `outstanding
>> partition`
>> > >> > (partition-x and partition-y) whose hashcode is 0x03 and 0x07, this
>> > >> > algorithm is going to split bundle the bundle into five new
>> bundles,
>> > >> > 0x00_0x03, 0x03_0x04( for parition-x), 0x04_0x07, 0x07_0x08( for
>> > >> > partition-y), 0x08_0x10.
>> > >>
>> > >> I think this approach have more potential with abnormal topic
>> isolation.
>> > >> If we can introduce
>> > >> some kind of bundle isolation strategy, (like broker-bundle affinity
>> and
>> > >> anti-affinity mechanism), we can easily isolate some unexpected
>> traffic to
>> > >> some empty brokers.
>> > >> IMO, this would improve the stability of broker cluster.
>> > >>
>> > >> Thanks,
>> > >> Haiting
>> > >>
>> > >> On 2022/02/17 15:47:15 Aloys Zhang wrote:
>> > >> > Hi Pulsar Community,
>> > >> >
>> > >> > This is a PIP discussion on how to support split partitions
>> belonging to
>> > >> > specified topics in a bundle.
>> > >> >
>> > >> > The issue can be found:
>> https://github.com/apache/pulsar/issues/13761
>> > >> >
>> > >> > I copy the content here for convenience, any suggestions are
>> welcome and
>> > >> > appreciated.
>> > >> >
>> > >> >
>> > >> > ## Motivation
>> > >> >
>> > >> > As we all know, a namespace bundle may contain lots of partitions
>> > >> belonging
>> > >> > to different topics.
>> > >> > The throughput of these topics may vary greatly. Some topics may
>> with
>> > >> very
>> > >> > high rate/throughput while other topics have a very low
>> rate/throughput.
>> > >> >
>> > >> > These partitions with high rate/throughput can cause broker
>> overload and
>> > >> > bundle unloading.
>> > >> > At this point, if we split bundle manually with
>> `range_equally_divide`
>> > >> or
>> > >> > `topic_count_equally_divide` split algorithm, there may need many
>> times
>> > >> > split before these high rate/through partitions assigned to
>> different
>> > >> new
>> > >> > bundles.
>> > >> >
>> > >> > For convenience, we call these high throughput topics `outstanding
>> > >> topic`
>> > >> > and their partitions `outstanding partition` in this PIP.
>> > >> >
>> > >> > ## Goal
>> > >> >
>> > >> > Our goal is to make it easier to split `outstanding partition`
>> into new
>> > >> > bundles.
>> > >> >
>> > >> > There are two alternative ways to achieve this. Either of them
>> will add
>> > >> a
>> > >> > new algorithm for bundle split. The difference is how the new
>> bundle
>> > >> split
>> > >> > algorithm is implemented.
>> > >> >
>> > >> > One algorithm is to split bundle by `outstanding topic` which will
>> split
>> > >> > the bundle into two new bundles and each new bundle contains an
>> equally
>> > >> > `outstanding partition` once a time.
>> > >> > E.g, a bundle contains lots of topic partitions, and only one
>> > >> `outstanding
>> > >> > topic`(T) with 2 `outstanding partition` (T-partition-n,
>> > >> Tpartition-n+1).
>> > >> > This algorithm split this bundle at the middle point of these two
>> > >> > partition's hashcode. This algorithm has a disadvantage, it can
>> only
>> > >> deal
>> > >> > with one `outstanding topic`.
>> > >> >
>> > >> > So we raised up another algorithm.
>> > >> >
>> > >> > The other algorithm is to split the bundle at the hashcode point
>> of the
>> > >> > `outstanding partition` which will split the bundle into three
>> bundles
>> > >> once
>> > >> > a time. The middle one contains the only point the hashcode of the
>> > >> > `outstanding partition, the left one is less than the hashcode, the
>> > >> right
>> > >> > one is more than the hashcode.
>> > >> > E.g. if we have a bundle 0x00_0x10 contains two `outstanding
>> partition`
>> > >> > (partition-x and partition-y) whose hashcode is 0x03 and 0x07, this
>> > >> > algorithm is going to split bundle the bundle into five new
>> bundles,
>> > >> > 0x00_0x03, 0x03_0x04( for parition-x), 0x04_0x07, 0x07_0x08( for
>> > >> > partition-y), 0x08_0x10.
>> > >> >
>> > >> > ## API Changes
>> > >> >
>> > >> > The Admin CLI `bin/pulsar-admin namespaces split-bundle -b
>> > >> ${bundle_range}`
>> > >> > will add a new parameter "--topic" or "-t" for `outstanding topic`
>> > >> name.
>> > >> >
>> > >> > The split interface changed from
>> > >> >
>> > >> > ```JAVA
>> > >> > void splitNamespaceBundle(String namespace, String bundle, boolean
>> > >> > unloadSplitBundles, String splitAlgorithmName)throws
>> > >> PulsarAdminException;
>> > >> > ```
>> > >> >
>> > >> > to
>> > >> >
>> > >> > ```java
>> > >> > void splitNamespaceBundle(String namespace, String bundle, boolean
>> > >> > unloadSplitBundles,
>> > >> > String splitAlgorithmName, String
>> topic)
>> > >> > throws PulsarAdminException;
>> > >> > ```
>> > >> >
>> > >> > ## Implementation
>> > >> >
>> > >> > There are changes both from the Admin CLI and the broker side.
>> > >> >
>> > >> > First, Admin CLI for split bundle should support to specify the
>> > >> > `outstanding topic`,
>> > >> >
>> > >> > ```java
>> > >> > /**
>> > >> > * Split namespace bundle.
>> > >> > *
>> > >> > * @param namespace
>> > >> > * @param bundle range of bundle to split
>> > >> > * @param unloadSplitBundles
>> > >> > * @param splitAlgorithmName
>> > >> > * @param topic
>> > >> > * @throws PulsarAdminException
>> > >> > */
>> > >> > void splitNamespaceBundle(String namespace, String bundle,
>> boolean
>> > >> > unloadSplitBundles,
>> > >> > String splitAlgorithmName, String
>> topic)
>> > >> > throws PulsarAdminException;
>> > >> >
>> > >> > ```
>> > >> >
>> > >> > ```java
>> > >> > /**
>> > >> > * Split namespace bundle asynchronously.
>> > >> > *
>> > >> > * @param namespace
>> > >> > * @param bundle range of bundle to split
>> > >> > * @param unloadSplitBundles
>> > >> > * @param splitAlgorithmName
>> > >> > */
>> > >> > CompletableFuture<Void> splitNamespaceBundleAsync(
>> > >> > String namespace, String bundle, boolean
>> unloadSplitBundles,
>> > >> > String splitAlgorithmName, String topic);
>> > >> > ```
>> > >> >
>> > >> > And for the broker side, first encapsulates the parameters for
>> bundle
>> > >> split
>> > >> > into a new class `BundleSplitOption`
>> > >> >
>> > >> > ```java
>> > >> > public class BundleSplitOption {
>> > >> > private NamespaceService service;
>> > >> > private NamespaceBundle bundle;
>> > >> > private String topic;
>> > >> > }
>> > >> > ```
>> > >> >
>> > >> > add a new split algorithm
>> > >> >
>> > >> > ```java
>> > >> > ublic class SpecifiedTopicCountEquallyDivideBundleSplitAlgorithm
>> > >> implements
>> > >> > NamespaceBundleSplitAlgorithm {
>> > >> > @Override
>> > >> > public CompletableFuture<List<Long>>
>> > >> getSplitBoundary(BundleSplitOption
>> > >> > bundleSplitOption) {
>> > >> >
>> > >> > });
>> > >> > }
>> > >> > }
>> > >> > ```
>> > >> >
>> > >> > add the new algorithm to `NamespaceBundleSplitAlgorithm`
>> > >> >
>> > >> > ```JAVA
>> > >> > String SPECIFIED_TOPIC_COUNT_EQUALLY_DIVIDE =
>> > >> > "specified_topic_count_equally_divide";
>> > >> >
>> > >> > List<String> AVAILABLE_ALGORITHMS =
>> > >> > Lists.newArrayList(RANGE_EQUALLY_DIVIDE_NAME,
>> > >> > TOPIC_COUNT_EQUALLY_DIVIDE,
>> > >> > SPECIFIED_TOPIC_COUNT_EQUALLY_DIVIDE);
>> > >> >
>> > >> > NamespaceBundleSplitAlgorithm
>> > >> SPECIFIED_TOPIC_COUNT_EQUALLY_DIVIDE_ALGO =
>> > >> > new
>> SpecifiedTopicCountEquallyDivideBundleSplitAlgorithm();
>> > >> > ```
>> > >> >
>> > >> > modify the `splitAndOwnBundle` and `splitAndOwnBundleOnceAndRetry`
>> for
>> > >> > [[NamespaceService.java](
>> > >> >
>> > >>
>> https://github.com/apache/pulsar/pull/13796/files#diff-99867fd2e328e0a689daa8f60d174aa96d87dd0d0fd29b9e68fe2e1f377b55a1)](https://github.com/apache/pulsar/pull/13796/files#diff-99867fd2e328e0a689daa8f60d174aa96d87dd0d0fd29b9e68fe2e1f377b55a1
>> > >> )
>> > >> >
>> > >> >
>> > >> > ```java
>> > >> > public CompletableFuture<Void> splitAndOwnBundle(NamespaceBundle
>> bundle,
>> > >> > boolean unload,
>> > >> >
>> > >> > NamespaceBundleSplitAlgorithm splitAlgorithm, String topic) {
>> > >> >
>> > >> > final CompletableFuture<Void> unloadFuture = new
>> > >> > CompletableFuture<>();
>> > >> > final AtomicInteger counter = new
>> > >> > AtomicInteger(BUNDLE_SPLIT_RETRY_LIMIT);
>> > >> > splitAndOwnBundleOnceAndRetry(bundle, unload, counter,
>> > >> > unloadFuture, splitAlgorithm, topic);
>> > >> >
>> > >> > return unloadFuture;
>> > >> > }
>> > >> > ```
>> > >> >
>> > >> > ```java
>> > >> > void splitAndOwnBundleOnceAndRetry(NamespaceBundle bundle,
>> > >> > boolean unload,
>> > >> > AtomicInteger counter,
>> > >> > CompletableFuture<Void>
>> > >> > completionFuture,
>> > >> >
>> NamespaceBundleSplitAlgorithm
>> > >> > splitAlgorithm,
>> > >> > String topic) {
>> > >> > ```
>> > >> >
>> > >> > Also, we change the REST api and broker.conf
>> > >> >
>> > >> > ```java
>> > >> > public void splitNamespaceBundle(
>> > >> > @Suspended final AsyncResponse asyncResponse,
>> > >> > @PathParam("property") String property,
>> > >> > @PathParam("cluster") String cluster,
>> > >> > @PathParam("namespace") String namespace,
>> > >> > @PathParam("bundle") String bundleRange,
>> > >> > @QueryParam("authoritative") @DefaultValue("false")
>> boolean
>> > >> > authoritative,
>> > >> > @QueryParam("unload") @DefaultValue("false") boolean
>> unload,
>> > >> > @QueryParam("topic") @DefaultValue("") String topic) {}
>> > >> > ```
>> > >> >
>> > >> > ```shell
>> > >> >
>> > >>
>> supportedNamespaceBundleSplitAlgorithms=range_equally_divide,topic_count_equally_divide,specified_topic_count_equally_divide
>> > >> > ```
>> > >> >
>> > >>
>> > >
>> >
>>
>
