This fixes the CollectionsAPIDistributedZkTest.testCoresAreDistributedAcrossNodes test failures. Due to the various tests for exceptional conditions, there were times where the session was not released causing stale data to remain in the policy session cache.
This test too makes assumptions about how replicas are placed. In the legacy assignment strategy, the replica of a given collection are spread equally across all nodes but with the new policy based strategy, all cores across collections are spread out. Therefore the assumptions in this test were wrong. I've changed this test to use the legacy assignment policy because testing the autoAddReplicas feature doesn't have to depend on new replica assignment strategies. This change also fixes a bug in Assign which used "collection" key instead of "cluster" to figure out which strategy to use.
The testNonRetryableRequests test makes an assumption that a collection's replicas are equally distributed among all nodes but with the policy engine it is not true. Instead the policy engine spreads out the cores belonging to all collections equally among all nodes. This is fixed by only creating the collection needed by tests in this class just-in-time.
Previously, the maxShardsPerNode parameter was not allowed on collections when autoscaling policy was configured. Also if an autoscaling policy was configured then the default was to set an unlimited maxShardsPerNode automatically. Now the maxShardsPerNode parameter is always allowed during collection creation and maxShardsPerNode should be set correctly (if required) regardless of whether autoscaling policies are in effect or not. The default value of maxShardsPerNode continues to be 1 as before. It can be set to -1 during collection creation to fall back to the old behavior of unlimited maxShardsPerNode when using autoscaling policy. This patch also fixes PolicyHelper to find the free disk space requirements of a new replica from the leader only if said leader node is alive.
ConcurrentUpdateSolrClient can batch together many documents when making
an indexing request to Solr. When adding an update request to the
current batch being made, it checks that the query-parameters of the
docs being added match those already in the batch. But prior to this
commit it never checked that the collections/cores were the same.
This could result in documents being sent to the wrong collection if the
same client is used to index documents to two different
cores/collections simultaneously.
This commit addresses this problem, ensuring that documents aren't added
to a batch directed at a different core/collection.
Wrapping is now a bit more lightweight: Does not create FieldInfo for fields that
can't be uninverted (saves mem) and can avoid wrapping the reader altogether if there's nothing to uninvert.
IndexSchema.getUninversionMap refactored to getUninversionMapper and no longer merges FieldInfos.
The cluster wide defaults structure has changed from {collectionDefaults: {nrtReplicas : 2}} to {defaults : {collection : {nrtReplicas : 2}}}. The old format continues to be supported and can be read from ZK as well as written using the V2 set-obj-property syntax but it is deprecated and will be removed in Solr 9. We recommend that users change their API calls to use the new format going forward.
This commit deprecates the min_rf parameter. Solr now always includes the achieved replication
factor in the update requests (as if min_rf was always specified). Also, reverts the changes
introduced in SOLR-8034, replicas that don't ack an update will have to recover to prevent
inconsistent shards.
Dawid Weiss
iverase
Erick Erickson
Alan Woodward
Jason Gerlowski
Christine Poerschke
Joel Bernstein
Jim Ferenczi
Shalin Shekhar Mangar
Jan Høydahl
Nicholas Knize
Noble Paul
Steve Rowe
Andrzej Bialecki
Cao Manh Dat
David Smiley
Tomas Fernandez Lobbe