Skip to content

Contrast vs OpenSearch

OpenSearch stores each span as a document on a cluster you run, fed by an OpenTelemetry Collector and Data Prepper. With Contrast you point an exporter at one endpoint. What happened after a tool error reads straight off each run.

Updated November 2026

At a glance

7.8×
faster than ClickHouse
8.1×
faster than Snowflake
90%
lower cost than Snowflake

Recorded races, October 2026. OpenSearch wasn't in the test.

How you're billed

List prices as each vendor publishes them, checked November 2026.

Contrast

Billed on
Credits, data scanned and storage
Price
$10 per TB scanned

Pay as you go. No cluster or pipeline to run.

OpenSearch

Billed on
Running nodes
Price
Your own nodes

Or instance hours or OCUs on Amazon OpenSearch Service.

Running a search cluster just to read your agents’ traces?

Agent questions span many runs and depend on order. Here's what each side takes to answer them.

Contrast

An owner starts TraceDB once and creates an API key. Your exporter sends OTLP to api.embrasure.ai/v1/traces, and Contrast runs everything behind it.

OpenSearch

OpenSearch's agent monitoring docs start with four parts: a cluster, Dashboards with an Observability workspace, a Collector and Data Prepper.

What happens in the two steps after a tool error?

PPL returns every span of each failing run and leaves working out what followed to you. TraceDB's paths() returns the most common two-step paths after a tool error in one query.

Contrast (TraceDB SQL)
SELECT path, count, entities
FROM paths(from => 'tool_error', steps => 2)
ORDER BY count DESC
LIMIT 10
OpenSearch (PPL)
source=otel-v1-apm-span-*
| where traceId in [
    source=otel-v1-apm-span-*
    | where like(name, 'execute_tool%') and `status.code` = 2
    | fields traceId
  ]
| sort traceId, startTime
| fields traceId, startTime, name, `status.code`

Contrast and OpenSearch, feature by feature

ContrastOpenSearch

Setup and ingestion

Nothing to deploy or operateContrastYes. Contrast runs TraceDB; an owner starts it.OpenSearchPartly. Cluster, Dashboards, Collector and Data Prepper.
OTLP over HTTP straight from any SDKContrastYes. Protobuf or JSON, gzipped or not.OpenSearchPartly. gRPC; unframed HTTP is opt-in.
OpenTelemetry GenAI conventionsContrastYes. Plus AI SDK, OpenInference and OpenLLMetry names.OpenSearchYes. Normalized by the Collector.

Storage and retention

Each run stored together in time orderContrastYes. One ordered run per trace.OpenSearchPartly. One document per span.
Retention without lifecycle policiesContrastYes. Each trace stays a run.OpenSearchPartly. Index State Management policies you write.
Durable copies in a second regionContrastYes. Versioned S3, replicated across regions.OpenSearchPartly. Snapshots you schedule through ISM.

Queries

SQL over tracesContrastYes. Standard SQL plus trace functions.OpenSearchYes. SQL and PPL through the SQL plugin.
Steps after an event in one functionContrastYes. paths(from => 'tool_error', steps => 2).OpenSearchPartly. Sort each trace's spans, read what follows.
Time-window checks between eventsContrastYes. followed_by() and preceded_by().OpenSearchPartly. Subqueries and joins over span documents.

Agents and pricing

MCP tools that query tracesContrastYes. Describe, query and run SQL on TraceDB.OpenSearchYes. The MCP Server runs search queries.
Traces beside your warehouse tablesContrastYes. TraceDB follows tables your definitions name.OpenSearchPartly. Only data indexed into the cluster.
No charge for running nodesContrastYes. Credits, data scanned and storage.OpenSearchNo. Your nodes, or AWS instance hours.

Running alongside OpenSearch

  1. In Contrast, open Sources, add Agent traces and have a workspace owner or admin start TraceDB.
  2. Create an API key on the same page. Contrast shows the OTEL_EXPORTER_OTLP_TRACES variables to set, or hands the setup to a coding agent.
  3. In the Collector that feeds Data Prepper, add an otlphttp exporter for Contrast to the same traces pipeline, so both get every span.
  4. Ask about your runs in chat, or connect your agents to the Contrast MCP server.

More comparisons

Speed figures are from TraceDB races recorded October 27, 2026 on a six-question investigation over 100 million synthetic events, median of 5 races after a warm-up. TraceDB and ClickHouse each ran on 4 vCPUs of one 8-vCPU VM in AWS us-east-1; Snowflake ran on an X-Small warehouse in AWS us-west-2. OpenSearch was not part of that test. The cost figure is the one the Contrast home page states for TraceDB against Snowflake. OpenSearch details are from its public documentation, checked November 2026: Agent monitoring, OTel trace source, Index State Management, PPL, OpenSearch MCP Server, FAQ, Amazon OpenSearch Service pricing. Product names are trademarks of their owners.