Hello Alex,
Thanks for the suggestion. I will try with both options. Most likely I will
use the custom AffinityBackupFilter implementation you provided and will
try to enhance zone skewness configurable.
Regards,
Amit
On Thu, Sep 19, 2024 at 3:16 PM Alex Plehanov <plehanov.a...@gmail.com>
wrote:
> Hello Amit,
>
> You can use ClusterNodeAttributeAffinityBackupFilter and introduce
> some virtual zones. For example, if you have 5 nodes in zone 1 and 5
> nodes in zone 2, you can assign 'zone1' attribute value to 3 nodes
> from zone 1, assign 'zone2' attribute value to 3 nodes from zone 2,
> and assign 'zone3' attribute value to 4 remaining nodes (2 from zone 1
> and 2 from zone 2). In this case, there will be 3 copies of each
> partition, each partition will be in each virtual zone, but nodes in
> virtual 'zone3' will contain a little bit less partitions than nodes
> from 'zone1' and 'zone2'.
>
> Or you can create your own backup filter, allowing no more than two
> nodes for the same attribute value, for example like this:
>
> public class MyAffinityBackupFilter implements
> IgniteBiPredicate<ClusterNode, List<ClusterNode>> {
> private final String attrName;
>
> public MyAffinityBackupFilter(String attrName) {
> this.attrName = attrName;
> }
>
> @Override public boolean apply(ClusterNode candidate,
> List<ClusterNode> previouslySelected) {
> Set<Object> usedAttrs = new HashSet<>();
>
> for (ClusterNode node : previouslySelected) {
> if (Objects.equals(candidate.attribute(attrName),
> node.attribute(attrName)) &&
> !usedAttrs.add(candidate.attribute(attrName)))
> return false;
> }
>
> return true;
> }
> }
>
> In this case you can achieve a more even distribution.
>
> чт, 19 сент. 2024 г. в 16:58, Amit Jolly <amit.jo...@gmail.com>:
> >
> > Hi Pavel,
> >
> > Well based upon documentation of
> ClusterNodeAttributeAffinityBackupFilter.java class. It says "This
> implementation will discard backups rather than place multiple on the same
> set of nodes. This avoids trying to cram more data onto remaining nodes
> when some have failed." and i have verified the same by running a small
> test with three node cluster (one assigned with node attribute as
> AVAILABILITY_ZONE=ZONE1 and other two assigned node attribute
> AVAILABILITY_ZONE=ZONE2) , Created a cache with 2 backups and using
> ClusterNodeAttributeAffinityBackupFilter in RendezvousAffinityFunction as
> below. After that added an entry into the cache and verified the nodes
> count for both primary and backup using cache affinity function. It
> returned 2 instead of 3.
> >
> > ClusterNodeAttributeAffinityBackupFilter backupFilter = new
> ClusterNodeAttributeAffinityBackupFilter("AVAILABILITY_ZONE");
> > RendezvousAffinityFunction rendezvousAffinityFunction = new
> RendezvousAffinityFunction();
> > rendezvousAffinityFunction.setAffinityBackupFilter(backupFilter);
> >
> > CacheConfiguration<String, String> cacheConfiguration = new
> CacheConfiguration<>();
> > cacheConfiguration.setBackups(2);
> > cacheConfiguration.setAffinity(rendezvousAffinityFunction);
> >
> > IgniteCache<String, String> cache =
> ignite.getOrCreateCache(cacheConfiguration);
> > cache.put("1","1");
> > Collection<ClusterNode> nodes =
> ((Ignite)cache.unwrap(Ignite.class)).affinity(cache.getName()).mapKeyToPrimaryAndBackups("1");
> > assertEquals(3, nodes.size()); //This fails even though i have three
> nodes (1 with node attribute AVAILABILITY_ZONE="ZONE1" and other two with
> node attribute AVAILABILITY_ZONE="ZONE2")
> >
> > PS: I started three nodes with Custom cache configuration
> IgniteConfiguration.setUserAttributes
> >
> > Server Node1
> > =====================================
> > Map<String, String> userAttributes = new HashMap<>();
> > userAttributes.put("AVAILABILITY_ZONE", "ZONE1");
> > IgniteConfiguration cfg = new IgniteConfiguration();
> > cfg.setUserAttributes(userAttributes);
> > Ignition.start(cfg);
> >
> > Server Node2
> > =====================================
> > Map<String, String> userAttributes = new HashMap<>();
> > userAttributes.put("AVAILABILITY_ZONE", "ZONE2");
> > IgniteConfiguration cfg = new IgniteConfiguration();
> > cfg.setUserAttributes(userAttributes);
> > Ignition.start(cfg);
> >
> > Server Node3
> > =====================================
> > Map<String, String> userAttributes = new HashMap<>();
> > userAttributes.put("AVAILABILITY_ZONE", "ZONE2");
> > IgniteConfiguration cfg = new IgniteConfiguration();
> > cfg.setUserAttributes(userAttributes);
> > Ignition.start(cfg);
> >
> >
> >
> https://ignite.apache.org/releases/latest/javadoc/org/apache/ignite/cache/affinity/rendezvous/ClusterNodeAttributeAffinityBackupFilter.html
> >
> > Thanks,
> >
> > Amit Jolly
> >
> > On Thu, Sep 19, 2024 at 12:51 AM Pavel Tupitsyn <ptupit...@apache.org>
> wrote:
> >>
> >> Hi Amit,
> >>
> >> > if the backup count is let's say 2, Ignite won't create a second
> backup as there are not enough zones
> >> Not correct - Ignite will create backups anyway.
> >> - A backup is a copy of a partition on another node
> >> - With 2 backups every partition will have 3 copies (1 primary, 2
> backup), all on different nodes (since you have 10 nodes)
> >> - Use ClusterNodeAttributeAffinityBackupFilter to ensure that at least
> one of the copies is in a different AZ
> >>
> >> And that is enough for 3 copies.
> >>
> >> On Thu, Sep 19, 2024 at 12:10 AM Amit Jolly <amit.jo...@gmail.com>
> wrote:
> >>>
> >>> Hi Team
> >>>
> >>> We are planning to run 10 node Ignite clusters in AWS with 5 nodes
> each into two availability zones. Using Kubernetes
> topologyspreadconstraints we have made sure that no two Ignite pods are
> started on the same virtual machine/node/host.
> >>>
> >>> I understand with ClusterNodeAttributeAffinityBackupFilter i can force
> ignite to store the backup in a different zone if backup count is 1.
> >>>
> >>> But if the backup count is let's say 2, Ignite won't create a second
> backup as there are not enough zones.
> >>>
> >>> My question is if i have backup count 2, How can I use
> ClusterNodeAttributeAffinityBackupFilter or (custom AffinityBackupFilter)
> to have at least one backup in each zone and another backup anywhere else
> where available.
> >>>
> >>> I think in order to achieve what I am thinking somehow I need
> currentTopologySnapshot available in
> ClusterNodeAttributeAffinityBackupFilter or custom AffinityBackupFilter
> >>>
> >>> Thanks,
> >>>
> >>> Amit Jolly
>
