Skip to content
Branch: master
Find file History
adriancole Upgrades versions to latest and tweaks ES InitialEndpointSupplier (#2799

Notable, the latest Armeria update allows us to focus Elasticsearch
`InitialEndpointSupplier` code on parsing as opposed to composition.
Latest commit a762a03 Sep 13, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
src Upgrades versions to latest and tweaks ES InitialEndpointSupplier (#2799 Sep 13, 2019 Migrate storage IT to JUnit5 (#2705) Jul 24, 2019
pom.xml [maven-release-plugin] prepare for next development iteration Aug 24, 2019


This is a CQL-based Cassandra storage component, built upon the Zipkin v2 api and model. This uses Cassandra 3.11.3+ features, but is tested against the latest patch of Cassandra 3.11.

CassandraSpanStore.getDependencies() returns pre-aggregated dependency links (ex via zipkin-dependencies).

The implementation uses the Datastax Java Driver 3.x. includes defaults that will operate against a local Cassandra installation.


Since the underlying driver uses SLF4J, Zipkin's storage layer also uses this (note SLF4J is supported out-of-the-box with no configuration in zipkin-server).

Zipkin's storage layer logs to the category "", but you may wish to see the entire "zipkin2" when troubleshooting. Depending on details desired, the underlying driver's category "com.datastax.driver.core" at debug level may help.

If you just want to see queries and latency, set the category "com.datastax.driver.core.QueryLogger" to debug or trace. Trace level includes bound values.

See Logging Query Latencies for more details.


This module conditionally runs integration tests against a local Cassandra instance.

This starts a docker container or attempts to re-use an existing cassandra node running on localhost.

If you run tests via Maven or otherwise when Cassandra is not running, you'll notice tests are silently skipped.

Results :

Tests run: 62, Failures: 0, Errors: 0, Skipped: 48

This behaviour is intentional: We don't want to burden developers with installing and running all storage options to test unrelated change. That said, all integration tests run on pull request via Travis.

Running a single test

To run a single integration test, use the following syntax:

$ ./mvnw -Dit.test='ITCassandraStorage$ITSpanStore#getTraces_duration' -pl zipkin-storage/cassandra clean verify

Strict trace ID

By default, trace identifiers are written at the length received to indexes and span tables. This means if instrumentation downgraded a 128-bit trace ID to 64-bit, it will appear in a search as two traces. This situation is possible when using unmaintained or out-of-date trace instrumentation.

By setting strict trace ID to false, indexes only consider the right-most 16 chars, allowing mixed trace length lookup at a slight collision risk. Retrieval of the 32-character trace ID is retained by concatenating two columns in the span table like so:

trace_id            text, // when strictTraceId=false, only contains right-most 16 chars
trace_id_high       text, // when strictTraceId=false, contains left-most 16 chars if present

It is important to only set strict trace ID false during a transition and revert once complete, as data written during this period is less intuitive for those using CQL, and contains a small collision risk.


This component is tuned to help reduce the size of indexes needed to perform query operations. The most important aspects are described below. See CassandraStorage for details.

Autocomplete indexing

Redundant requests to store autocomplete values are ignored for an hour to reduce load. This is implemented by DelayLimiter

Trace indexing

Indexing in CQL is simplified by SASI, for example, reducing the number of tables from 7 down to 4 (from the original cassandra schema). SASI also moves some write-amplification from CassandraSpanConsumer into C*.

CassandraSpanConsumer directly writes to the tables span, trace_by_service_remote_service trace_by_service_span and span_by_service. The latter service based indexes amplify writes by a factor of the distinct service names (Span.localServiceName).

Other amplification happens internally to C*, visible in the increase write latency (although write latency remains performant at single digit milliseconds).

span indexing

When queries only include a time range, trace ids are returned from a ts_uuid range. This means no indexes are used when GET /api/v2/traces includes no parameters or only endTs or lookback.

Two secondary (SASI) indexes support annotationQuery with serviceName:

  • annotation_query supports LIKE (substring match) in ░error░error=500░
  • l_service in used in conjunction with annotation_query searches.

Ex, GET /api/v2/traces?serviceName=tweetiebird&annotationQuery=error results in a single trace ID query against the above two indexes.

Note: annotations with values longer than 256 characters are not written to the annotation_query SASI, as they aren't intended for use in user queries.

trace_by_service_X indexing

trace_by_service_X rows are answers to a shard of trace query. A query request is broken down into possibly multiple shards based on our index implementation.

Ex. GET /api/v2/traces?serviceName=tweetiebird%remoteService=s3

Breaks down into two query shards (this example omits time range and limit)

  • (service=tweetiebird, span=)
  • (service=tweetiebird, remote_service=s3)

The results intersect prioritizing on timestamp to return the distinct trace IDs needed for a follow-up fetch.

trace_by_service_remote_service indexing

For example, a span in trace ID 1 named "get" created by "tweetiebird", accessing the remote service "s3" results in the following row:

  • service=service1, span=remote_service, ts=timestamp_millis, trace_id=1

This index is only used when the remoteServiceName query is used. Ex.

  1. GET /api/v2/traces?serviceName=tweetiebird&remoteServiceName=s3
  2. GET /api/v2/traces?serviceName=tweetiebird&maxDuration=199500&remoteServiceName=s3

trace_by_service_span indexing

For example, a span in trace ID 1 named "get" created by "service1", taking 20 milliseconds results in the following rows:

  1. service=service1, span=get, trace_id=1, ts=timestamp_millis, duration=200
  2. service=service1, span=, trace_id=1, ts=timestamp_millis, duration=200

Here are corresponding queries that relate to the above rows:

  1. GET /api/v2/traces?serviceName=service1&spanName=get
  2. GET /api/v2/traces?serviceName=service1&spanName=get&minDuration=200000
  3. GET /api/v2/traces?serviceName=service1&minDuration=200000
  4. GET /api/v2/traces?spanName=get
  5. GET /api/v2/traces?maxDuration=199500

As you'll notice, the duration component is optional, and stored in millisecond resolution as opposed to microsecond (which the query represents). The final query shows that the input is rounded up to the nearest millisecond.

The reason we can query on duration is due to a SASI index. Eventhough the search granularity is millisecond, original duration data remains microsecond granularity. Meanwhile, write performance is dramatically better than writing discrete values, due to fewer distinct writes.

Disabling indexing

Indexing is a good default, but some sites who don't use Zipkin UI's "Find a Trace" screen may want to disable indexing. This means indexing schema won't be setup, nor written at runtime. This increases write throughput and reduces size on disk by not amplifying writes with index data.

Disabling search disables indexing.


Time-To-Live is default now at the table level. It can not be overridden in write requests.

There's a different default TTL for trace data and indexes, 7 days vs 3 days respectively. The impact is that you can retrieve a trace by ID for up to 7 days, but you can only search the last 3 days of traces (ex by service name).


Time-series data is compacted using TimeWindowCompactionStrategy, a known improved over DateTieredCompactionStrategy. Data is optimised for queries within a single day. The penalty of reading multiple days is small, a few disk seeks, compared to the otherwise overhead of reading a significantly larger amount of data.


Benchmarking the new datamodel demonstrates a significant performance improvement on reads. How much of this translates to the Zipkin UI is hard to tell due to the complexity of CassandraSpanConsumer and how searches are possible. Benchmarking stress profiles are found in traces-stress.yaml and trace_by_service_span-stress.yaml and span_by_service-stress.yaml.

You can’t perform that action at this time.