Skip to content

Latest commit

 

History

History
130 lines (86 loc) · 7.4 KB

README.md

File metadata and controls

130 lines (86 loc) · 7.4 KB
Raw

zipkin-storage-kafka design

Goals

  • Provide a fast and reliable storage that enable extensibility via Kafka topics.
  • Provide full storage functionality via streaming aggregations (e.g., spans queries, dependency graph).
  • Create a processing space where additional enrichment can be plugged-in into the tracing pipeline.
  • Remove need for additional storage when Kafka is available.

Kafka Zipkin Storage

Storage is composed by 4 main components:

  • Span Partitioning: repartition of collected span batches into spans keyed by traceId
  • Span Aggregation: processing of partitioned spans into aggregated traces and later into dependency links.
  • Trace and Dependency Storage: building local state stores to support search and query APIs for traces and dependencies.

Span Partitioning

This component processes span batches, received via HTTP, Kafka, ActiveMQ, etc.; take elements with same trace-id and keyed them to be stored on zipkin-spans topic for repartition.

Property Environment Variable Description
zipkin.storage.kafka.span-partitioning-enabled SPAN_PARTITIONING_ENABLED false disables span partitioning. Consider that Aggregation component requires partitioned spans. Defaults to true.
zipkin.storage.kafka.partitioned-spans-topic none Kafka topic where partitioned spans will be stored. Key type: trace-id string, Value type: list of spans with same trace-id. Defaults to zipkin-spans

This component is currently compensating how KafkaSender (part of Zipkin-Reporter) is reporting spans to Kafka, by grouping spans into batches and sending them to a un-keyed Kafka topic.

Source code: KafkaSpanConsumer.java

Span Aggregation

Partitioned spans are processed to produced two aggregated streams: trace and dependency.

Trace Stream

Spans are grouped by ID and stored on a local Session window, where the traceId becomes the token, and trace-timeout (default: 1 minute) (i.e. period of time without receiving a span with the same session; also known as session inactivity gap in Kafka Streams) defines if a trace is still active or not. This is evaluated on the next span received on the stream-- regardless of incoming traceId. If session window is closed, a trace message is emitted to the traces topic.

Session Windows

Each color represents a trace. The longer trace timeout we have, the longer we wait to close a window and the longer we wait to emit traces downstream for dependency link and additional aggregations; but also the more consistent the trace aggregation is. If we choose a smaller gap, then we emit traces faster with the risk of breaking traces into smaller chunks, and potentially affecting counters downstream.

Dependency Stream

Once traces are emitted downstream as part of the initial processing, dependency links are evaluated on each trace, and emitted the dependencies topic for further metric aggregation.

Property Environment Variable Description
zipkin.storage.kafka.span-aggregation-enabled SPAN_AGGREGATION_ENABLED false disables span aggregation. Consider that Storage components will require traces and dependency streams as input. Defaults to true.
zipkin.storage.kafka.aggregation-spans-topic none Kafka topic where partitioned spans will be consumed from. Key type: trace-id string. Value type: binary list of spans. Defaults to zipkin-spans
zipkin.storage.kafka.aggregation-trace-topic none Kafka topic where aggregated traces will be stored. Key type: trace-id string. Value type: binary list of spans. Defaults to zipkin-trace
zipkin.storage.kafka.aggregation-dependency-topic none Kafka topic where aggregated dependencies will be stored. Key type: string with parent|child format. Value type: dependency link. Defaults to zipkin-dependency

Kafka Streams topology: trace aggregation

Trace Storage

This component build local stores from state received on spans Kafka topic for traces, service names and autocomplete tags.

This component supports search and query APIs on top of local state stores build by the Store Kafka Streams component.

Get Service Names/Get Span Names/Get Remote Service Names

These queries are supported by service names indexed stores built from spans Kafka topic.

Kafka Streams store names:

  • zipkin-service-names: key/value store with service name as key and value.
  • zipkin-span-names: key/value store with service name as key and span names list as value.
  • zipkin-remote-service-names: key/value store with service name as key and remote service names as value.

Get Trace/Find Traces

These queries are supported by two key value stores:

  • zipkin-traces: indexed by traceId, contains span list status received from spans Kafka topic.
  • zipkin-traces-by-timestamp: list of trace IDs indexed by timestamp.

GetTrace query is supported by zipkin-traces store. FindTraces query is supported by both: When receiving a query request time range is used to get trace IDs, and then query request is tested on each trace to build a response.

Get Keys/Get Values

Supported by a key-value containing list of values valid for autocompleteKeys.

  • zipkin-autocomplete-tags: key-value store.
Property Environment Variable Description
zipkin.storage.kafka.trace-by-id-query-enabled TRACE_BY_ID_QUERY_ENABLED false disables query by trace ID GET /traces/{trace_id}. Defaults to true.
zipkin.storage.kafka.trace-search-enabled TRACE_SEARCH_ENABLED false disables trace search functionality GET /traces and indexes: service names, span names, tags. Defaults to true.
zipkin.storage.kafka.storage-spans-topic none Kafka topic where trace spans are consumed from to build state store. Expected value: list of spans; it does not use record keys. Defaults to zipkin-spans

Source code: TraceStorageTopology

Kafka Streams topology: trace storage

Dependency Storage

This component build local state store from dependency stream. It builds a 1 minute time-window when counts calls and errors. When a request is received, time range is used to pick valid windows and join counters.

Windowed store: zipkin-dependencies.

Property Environment Variable Description
zipkin.storage.kafka.dependency-query-enabled DEPENDENCY_QUERY_ENABLED false disables dependency query store GET /dependencies. Consider that Aggregation component requires partitioned spans. Defaults to true.
zipkin.storage.kafka.storage-dependency-topic none Kafka topic where dependencies are consumed from to build state store. Defaults to zipkin-dependency

Kafka Streams topology: dependency storage

Source code: DependencyStorageTopology