Monitor SAI indexes
Both virtual tables and JMX-based metrics can be used to monitor the SAI indexes in your cluster.
SAI virtual tables
You can refer to data in the following Apache Cassandra virtual tables to determine the status of indexes created with SAI:
-
system_views.indexes
— provides information at the column index level, including the index name, number of indexed SSTables, disk usage, and index state. From the index state, the data reveals if the index is currently building, and whether the index can be queried.When you
DROP
/ recreate an SAI index, you are not blocked from entering queries that do not use the index. However, you cannot use that SAI index (based on the same column) until it has finished building and is queryable. To determine the current state of a given index, query thesystem_views.indexes
virtual table. Example:SELECT is_queryable,is_building FROM system_views.indexes WHERE keyspace_name='keyspace' AND table_name='table' AND index_name='index';
-
system_views.sstable_indexes
— describes individual SSTable indexes, and includes information around disk size, min/max row ID, the min/max ring token, and the write-time version of the index. -
system_views.sstable_index_segments
— describes the segments of the SSTable indexes. It exposes the segment row ID offset and most of the information in the SSTable-level virtual table, specifically at a segment granularity. For more details, refer to Virtual tables for SAI indexes and SSTables.
SAI tracing
SAI provides tracing capability just like other database components.
Information is captured by the system_traces
keyspace.
You can enable tracing in CQLSH with TRACING ON
, or in the Cassandra driver with statement.enableTracing()
.
The number of rows filtered by a particular query will show up in the CQL query trace. Example:
Index query accessed memtable indexes, 2 SSTable indexes, and 2 segments, post-filtered 14 rows in 14 partitions, and took 88582 microseconds.
For details about tracing, refer to CQL TRACING.
SAI metrics
SAI provides a number of metrics to help you monitor the health of your indexes.
The categorized data:
-
Global indexing metrics
-
Table query metrics
-
Per query metrics
-
Column query metrics per index
-
Range slice metrics
For example, you can use metrics to get the current count of total partition reads since the node started for cycling.cyclist_semi_pro
.
The keyspace and table were defined in SAI quickstart.
This metric’s ObjectName
: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=cycling,table=cyclist_semi_pro,scope=TableQueryMetrics,name=TotalPartitionReads
.
The metrics are exposed via JMX, so any JMX-based tool can be used to monitor.
Global indexing metrics
ObjectName: org.apache.cassandra.metrics,type=StorageAttachedIndex,name=<metric>
The global indexing metrics for this node are:
-
ColumnIndexBuildsInProgress
— The number of individual on-disk column indexes currently being built. -
SegmentBufferSpaceLimitBytes
— The limit on heap used to buffer SSTable index segments during compaction and index builds.In cassandra.yaml,
segment_write_buffer_space_mb
limits the amount of heap used to build on-disk column indexes during compaction and initial builds. The default is 1024 MB.
For example, if there is only one column index building, SAI can buffer up to segment_write_buffer_space_mb
.
If there is one column index building per table across 8 compactors, each index will be eligible to flush once it reaches (segment_write_buffer_space_mb / 8
) MBs.
-
SegmentBufferSpaceUsedBytes
— The heap currently being used to buffer SSTable index segments during compaction and index builds.
At any given time, the minimum size for a flushing segment, in bytes, is ( |
Index group metrics
ObjectName: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=<keyspace>,table=<table>,scope=IndexGroupMetrics,name=<metric>
The index group metrics for the given keyspace and table:
-
DiskUsedBytes
— Size in bytes on disk for the given table’s SAI indices. -
IndexFileCacheBytes
— Size in bytes of memory used by the on-disk data structure of the per-column indices. -
OpenIndexFiles
— Number of open index files for the given table’s SAI indices.
Per query metrics
ObjectName: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=<keyspace>,table=<table>,scope=PerQuery,name=<metric>
The per query metrics for the given keyspace and table include:
-
RowsFiltered
— A histogram of the number of rows post-filtered per query since the node started. -
QueryLatency
— Overall query latency percentiles (in microseconds) and one/five/fifteen minute query throughput. -
PartitionReads
— Histogram over the number of partitions read per query. -
SSTableIndexesHit
— Histogram over the number of SSTable indexes read per query. -
KDTreeChunkCacheLookups
— Histogram over the number of chunk cache lookups while reading kd-tree index files per query. -
KDTreeChunkCacheMisses
— Histogram over the number of chunk cache misses while reading kd-tree index files per query.
Table query metrics
ObjectName: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=<keyspace>,table=<table>,scope=TableQueryMetrics,name=<metric>
The table query metrics for the given keyspace and table:
-
TotalPartitionReads
— Total partition reads by all queries since the node started. -
TotalQueriesCompleted
— Total number of successfully completed queries since the node started. -
TotalQueryTimeouts
— Total number of timeouts from queries since the node started. -
TotalRowsFiltered
— Total number of rows post-filtered by all queries since the node started.
Table state metrics
ObjectName: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=<keyspace>,table=<table>,scope=TableStateMetrics,name=<metric>
The table state metrics for the given keyspace and table:
-
DiskPercentageOfBaseTable
— SAI size on Disk as a percentage of table size per table. -
DiskUsedBytes
— Size on-disk in bytes of SAI indices per table. -
TotalIndexBuildsInProgress
— Status of SAI indices per table currently in theis_building
state. -
TotalIndexCount
— Total number of SAI indices per table. -
TotalQueryableIndexCount
— Status of SAI indices per table currently in theis_querable
state.
Column query metrics for each numeric index
ObjectName: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=<keyspace>,table=<table>,index=<index>,scope=ColumnQueryMetrics,name=<metric>
The column query metrics for the given keyspace, table, and index include:
-
KDTreeNiceTryLatency
— For numeric indexes, such asage_sai_idx
in the quickstart examples, this metric may be used to present a histogram of the times spent waiting for chunk cache misses during kd-tree intersection (in microseconds) and one/five/fifteen minute chunk miss throughputs.The throughputs are zero if there are no cache misses during kd-tree intersection.
Column query metrics for each string index
ObjectName: org.apache.cassandra.metrics:type=StorageAttachedIndex,keyspace=<keyspace>,table=<table>,index=<index>,scope=ColumnQueryMetrics,name=<metric>
The column query metrics for the given keyspace, table, and index include:
-
TermsLookupLatency
— For string indexes, such ascountry_sai_idx
in the quickstart examples, this metric shows terms lookup latency percentiles (in microseconds) per one/five/fifteen minute query throughput.
Range slice metrics
ObjectName: org.apache.cassandra.metrics:type=ClientRequest,scope=RangeSlice,name=<metric>
The RoundTripsPerReadHistogram
metric tracks the number of round-trip requests sent for range query commands from the coordinator.
Fewer requests typically mean the server is operating more efficiently than ones requiring more requests to satisfy the same range queries.
Latency metric tracks the min, max, mean as well as a set of percentiles for range read requests latency. Timeouts metric tracks the number of timeouts for range read requests.