Vector index metrics

This page lists the metrics related to the vector indexes that you generate in AlloyDB for PostgreSQL. You can view these metrics using the pg_stat_ann_indexes view that is available when you install the alloydb_scann extension.

For more information about viewing the metrics, see View vector index metrics.

Usability metrics

The usability metrics include metrics that help you understand the state of index utilization with metrics such as, index configuration and number of index scans.

Metric name	Data type	Description
`relid`	`OID`	Unique identifier of the table that contains the vector index.
`indexrelid`	`OID`	Unique identifier of the vector index.
`schemaname`	`NAME`	Name of the schema to which index belongs.
`relname`	`NAME`	Name of the table that contains the index.
`indexrelname`	`NAME`	Name of the index.
`indextype`	`NAME`	Type of the index. This value is always set to `scann`.
`indexconfig`	`TEXT[]`	Configuration, such as leaves count and quantizer, defined for the index when it was created.
`indexsize`	`TEXT`	Size of the index.
`indexscan`	`BIGINT`	Number of index scans initiated on the index.

Tuning metrics

Tuning metrics provide insights into your current index optimization, allowing you to apply recommendations for faster query performance.

Metric name	Data type	Description
`insertcount`	`BIGINT`	Number of insert operations on the index. This metric also includes any number of rows that existed before the index was created.
`updatecount`	`BIGINT`	Number of update operations on the index. This metric doesn't take into account any HOT updates.
`deletecount`	`BIGINT`	Number of delete operations on the index.
`distribution`	`JSONB`	Vector distributions across all partitions for the index. The following fields show the distribution: `maximum (INT8)`: Maximum number of vectors across all partitions. `minimum (INT8)`: Minimum number of vectors across all partitions. `average (FLOAT)` : Average number of vectors across all partitions. `outliers (INT8[])`: Top outliers across all partitions. This value shows the top 20 outliers. Note: Due to the inherent characteristics of the K-means clustering algorithm, there will always be some degree of variance in the distribution of vectors across partitions, even when the index is initially created.
`distributionpercentile`	`JSONB`	Vector index distribution (in Preview) helps you understand the distribution of vectors between partitions and `num_leaves` of your ScaNN index. Vector index distribution contains buckets for 10, 25, 50, 75, 90, 95, 99, and 100th percentiles. Each bucket contains the following values: Number of vectors present in the partition at the specified percentile. Number of partitions between the previous percentile and the current percentile buckets. Note: Due to the inherent characteristics of the K-means clustering algorithm, there is always some degree of variance in the distribution of vectors across partitions, even when the index is initially created.

Tuning recommendation based on the metrics

Mutation

The insertcount, updatecount, and deletecount metrics together show the changes or mutations in the vector for the index.

The index is created with a specific number of vectors and partitions. When operations such as insert, update, or delete are performed on the vector index, it only affects the initial set of partitions where the vectors reside. Consequently, the number of vectors in each partition fluctuates over time, potentially impacting recall, QPS, or both.

If you encounter slowness or accuracy issues such as low QPS or poor recall, in your ANN search queries over time, then consider reviewing these metrics. A high number of mutations relative to the total number of vectors could indicate the need for reindexing.

Distribution

The distribution metric shows the vector distributions across all partitions.

When you create an index, it is created with a specific number of vectors and fixed partitions. The partitioning process and subsequent distribution occurs based on this consideration. If additional vectors are added, they are partitioned among the existing partitions, resulting in a different distribution compared to the distribution when the index was created. Since the final distribution does not consider all vectors simultaneously, the recall, QPS, or both might be affected.

If you observe a gradual decline in the performance of your ANN search queries, such as slower response times or reduced accuracy in the results (measured by QPS or recall), then consider checking this metric and reindexing.

Distribution percentile

The distributionpercentile metric ([Preview](https://cloud.google.com/products#product-launch-stages)), is a vector index distribution in the pg_stat_ann_indexes view that helps you understand the distribution of vectors between partitions and num_leaves of your ScaNN index.

When you create an alloydb_scann index on the initial set of rows by setting num_leaves, the index can change the distribution of vectors across the partitions due to data operations (skew mutations), or the number of vectors might increase significantly. These changes can lead to degradation of QPS, recall, or both. The vector index distribution can give you signals if the mutation causes a change in the index distribution. This information can help you determine if a reindex is required, or if a change in search time configurations can help improve query performance.

Usually, vector distribution isn't uniform across different partitions in the index. This is called a non-uniform distribution and it doesn't require `REINDEX`. A non-uniform distribution has the following characteristics:

The variance of the number of vectors is low. Variance can be calculated as

$(P100(num\_vectors) - p10(num\_vectors))*(\frac{num\_leaves}{total\_num\_row})$
The number of partitions with 0 vectors is low, and might be less than 30% of partitions.
The variance of the number of partitions is low.
$ variance _{p} = abs(p_{num\_partitions} - num\_leaves * (p_{percentile} - p-1_{percentile})) $ where "p" is a vector index distribution bucket.
The number of vectors at any percentile is

$< 8 x (\frac{num\_rows\ during\ index\ creation\ time}{ num\_leaves})$

When these conditions aren't satisfied, REINDEX might be required based on how much QPS and recall are affected.

The following scenarios, while less common than non-uniform distribution, can occur:

Approximate Uniform Index: When most of the partitions have the same number of non-zero vectors and the variance of the number of vectors is low, it is an approximate uniform index. `REINDEX` is required if the number vectors in each partition are $> 8 * average vector$ at index_creation_time.
Sparse Index: A sparse index occurs when any of the conditions of a non-uniform index aren't fulfilled. REINDEX is required with a change of configurations.
A sparse index also occurs where > 50% of the partitions are empty. This scenario causes the vectors to be concentrated in a small number of partitions, which increases the number of vectors in each partition. When this happens, drop the index and recreate it.