YSQL Distributed Tracing TECH PREVIEW
YSQL queries spend time in many places: parsing, planning, execution, transaction commit, and RPC calls to tablet servers. YugabyteDB can export timing data for these stages as OpenTelemetry (OTel) traces so you can inspect a waterfall view of query execution in tools such as Jaeger, Grafana Tempo, or Honeycomb.
Distributed Tracing is TP and available in v2025.2.4.0 and later, and is currently only available for YSQL. Tracing inside tablet servers is not included in this release.
How it works
YSQL Distributed Tracing follows the W3C Trace Context standard. Your application supplies a traceparent value. YugabyteDB creates spans for the query lifecycle and exports them to an OTel collector over OTLP/HTTP.
Each traced query produces a trace made up of spans. Spans are nested to show where time is spent. For example, planning, execution, commit, and individual RPC calls to PgClientService.
When tracing is disabled (the default), there is no measurable performance impact. When tracing is enabled for a query, other queries and other YSQL backends are not affected.
Configure Distributed Tracing
To configure Distributed Tracing, set the following YB-TServer flags on each node in the cluster. Changing these flags requires a YB-TServer restart.
| Flag | Description | Default |
|---|---|---|
| otel_collector_traces_endpoint | OTLP/HTTP URL where spans are exported. For example, http://<collector-host>:4318/v1/traces.Setting this flag enables tracing infrastructure in each YSQL backend process. |
Empty |
| otel_batch_max_queue_size | Maximum number of spans buffered in the queue. Spans beyond this limit are dropped. Must be greater than 0 and at least as large as otel_batch_max_export_batch_size. |
2048 |
| otel_batch_schedule_delay_ms | Milliseconds between batch exports. Lower values reduce export latency but increase export frequency. | 5000 |
| otel_batch_max_export_batch_size | Maximum spans per export batch. Must be greater than 0 and no larger than otel_batch_max_queue_size. |
512 |
Prerequisites
- YSQL must be enabled on the cluster.
- The OTLP/HTTP endpoint must be reachable from each YB-TServer node.
- Because Distributed Tracing is a preview feature, add
otel_collector_traces_endpointto the allowed_preview_flags_csv list before setting it.
Set up tracing
The following example uses Jaeger as the trace backend. Jaeger accepts OTLP over HTTP on port 4318.
Start Jaeger
Run Jaeger all-in-one with OTLP enabled:
docker run --rm --name jaeger \
-e COLLECTOR_OTLP_ENABLED=true \
-p 16686:16686 \
-p 4318:4318 \
jaegertracing/all-in-one:1.61
Open the Jaeger UI at http://localhost:16686.
For a local single-node cluster started with yugabyted:
./bin/yugabyted start \
--tserver_flags "allowed_preview_flags_csv=otel_collector_traces_endpoint,otel_collector_traces_endpoint=http://127.0.0.1:4318/v1/traces"
If you use YugabyteDB Anywhere, set the flags using Edit configuration flags.
Note
Ifotel_collector_traces_endpoint is not set, attempting to use the yb_dist_tracecontext configuration parameter returns an error indicating that Distributed Tracing is not enabled.
Trace a query
After the cluster is running with tracing configured, enable tracing for individual queries using one of the following methods.
SQL comment (per query)
Prepend or append a block comment that contains a W3C traceparent value. The comment must be the first block comment at the start of the query, or the last block comment at the end of the query. If the comment is not the first or last, the traceparent is not parsed.
/*traceparent='00-0af7651916cd43dd8448eb211c80319c-b7ad6b7169203331-01'*/
SELECT * FROM orders WHERE customer_id = 500;
You can also place the comment at the end of the statement:
SELECT * FROM orders WHERE customer_id = 500
/*traceparent='00-0af7651916cd43dd8448eb211c80319c-b7ad6b7169203331-01'*/;
Configuration parameter (per session or transaction)
Set the yb_dist_tracecontext YSQL configuration parameter.
For example, to trace every query in the current transaction:
BEGIN;
SET LOCAL yb_dist_tracecontext = 'traceparent=''00-00000000000000000000000000000001-0000000000000005-01''';
SELECT * FROM users WHERE id = 10;
UPDATE accounts SET balance = balance - 100 WHERE id = 10;
COMMIT;
To stop tracing for the session:
RESET yb_dist_tracecontext;
Parameter takes precedence
If you provide both atraceparent parameter and SQL comment, the parameter takes priority and a warning is emitted.
View traces
Run a traced query, wait a few seconds for spans to be batched and exported (controlled by otel_batch_schedule_delay_ms), then open the Jaeger UI.
- Select Service
ysql. - Click Find Traces.
- Open a trace to see spans such as
query,parse,plan,execute,commit, and RPC spans.
Trace data
Process attributes
Each YSQL backend exports process metadata with every trace:
| Attribute | Description |
|---|---|
| service.name | Always ysql. |
| service.instance.id | UUID of the YB-TServer node running the YSQL backend. |
| process.pid | Operating-system PID of the YSQL backend process. |
Span attributes
Every span includes standard OpenTelemetry fields such as operation name, span ID, parent span ID, trace ID, start time, duration, span kind, and status.
Additional attributes depend on the span type:
| Span | Additional attributes |
|---|---|
Root (query) |
query.text (truncated to 256 characters), user.id (PostgreSQL role OID) |
RPC (for example, rpc yb.tserver.PgClientService.Perform) |
rpc.table_names — tables accessed by the RPC |
Typical span names include:
- Query lifecycle:
query,parse,rewrite,plan,execute,commit,abort - Extended query protocol:
ext.parse,ext.bind,ext.execute,ext.sync,ext.describe,ext.flush - RPC calls:
rpc yb.tserver.PgClientService.Performand related RPC operation names
Join application traces
Because YugabyteDB accepts W3C traceparent values, you can continue a trace started in application code. Pass the same traceparent in a SQL comment or yb_dist_tracecontext parameter so database spans appear as children of your application's root span in your observability backend.
Related topics
- Monitor with Active Session History - sample-based view of database wait events
- Query tuning - optimize query performance with EXPLAIN, pg_stat_statements, and related tools
- Jaeger integration - use YCQL as Jaeger trace storage (separate from YSQL query tracing)