In the Vald cluster, the same Index is replicated to multiple agents due to the
index_replica setting. However, inconsistencies between replicas may occur due to pod eviction or the occurrence of OOM killer during vector insertions. For example,
- The timestamp of the index differs between agents (some agents have an old index saved and it has not been updated).
- The number of replicas does not meet the value set in
To resolve these inconsistencies, you can use the
Index Correction feature.
Index Correction is implemented as a
CronJob, checking the consistency between replicas regularly and resolving any inconsistencies.
Turns the index correction feature on/off.
Sets the interval for the job start in cron notation (the default value is
3 6 * * *, which means 3:06 AM every day).
Temporary suspension setting for CronJob.
schedule: "3 6 * * *"
Under conditions of 10 million identical vectors(not including
index_replica) and 10 agent replicas, the processing takes about 30~40 minutes (this is only a reference, and the actual execution time may vary depending on the infrastructure). Time complexity of the process is
O(MN)where M is the number of identical vector items and N is the number of agent replicas.
index_replicadoes not matter for the processing time.
Forbidis set internally, so a new job will not be created while an existing job is running. In other words, if the process does not finish within the interval specified by the schedule, the next job will not be scheduled.
Index operations during correction
Vector operations performed after the start of the index correction job are not considered in that job.