Skip to content

Contrast vs Datadog

Datadog bills agent traces per LLM span and keeps them for 15 days. Contrast keeps every trace you export as a run in TraceDB and bills for data scanned and storage, so what happened after an error reads straight off each run.

Updated November 2026

At a glance

7.8×
faster than ClickHouse
90%
lower cost than Snowflake
$0
per host or span

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

How you're billed

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

Contrast

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

50 GB of scans a month free. Hosts and spans don't set the price.

Datadog

Billed on
Hosts and spans
Price
Per host for APM

Plus spans past each host's allotment. Agent Observability bills per LLM span.

Paying per span for agent traces that expire in 15 days?

Agent questions cover thousands of runs, like what happens after a tool error. Here's how each side answers them.

Contrast

Every trace you export stays a run in your TraceDB database, durable in S3 once acknowledged.

Datadog

APM keeps the spans its retention filters index for 15 days. Agent Observability keeps traces for 15 days, or 90 with an add-on.

The two steps after a tool error

In DDSQL you pull every span, rebuild each trace's order with LEAD() and filter for tool errors afterwards. TraceDB stores each run in order, so paths() reads the next two steps directly.

Contrast
SELECT path, count, entities
FROM paths(from => 'tool_error', steps => 2)
ORDER BY count DESC
LIMIT 10
Datadog
SELECT next_1, next_2, count(*) AS errors
FROM (
  SELECT kind, status,
    LEAD(name, 1) OVER (PARTITION BY trace_id
      ORDER BY start_ns) AS next_1,
    LEAD(name, 2) OVER (PARTITION BY trace_id
      ORDER BY start_ns) AS next_2
  FROM dd.llm_observability(event_type => 'span',
    columns => ARRAY['@trace_id', '@name',
      '@meta.span.kind', 'status', '@start_ns'])
    AS (trace_id VARCHAR, name VARCHAR, kind VARCHAR,
      status VARCHAR, start_ns BIGINT)
) t
WHERE kind = 'tool' AND status = 'error'
GROUP BY next_1, next_2
ORDER BY errors DESC LIMIT 10

Contrast and Datadog, feature by feature

ContrastDatadog

Runs

Every exported trace kept as a runContrastYes. No retention filter decides what stays.DatadogPartly. Retention filters pick the indexed spans.
Spans stored in run orderContrastYes. Each run's spans stay in order.DatadogPartly. Order rebuilt per query.

Querying runs

SQL over agent tracesContrastYes. TraceDB SQL with trace functions.DatadogYes. DDSQL over LLM Observability data.
Path, funnel and session functionsContrastYes. paths(), funnel(), sessions() and followed_by().DatadogPartly. Window functions like LEAD() and LAG().

Instrumentation

OpenTelemetry over OTLP/HTTPContrastYes. Protobuf or JSON, from any SDK or Collector.DatadogYes. Agent, Collector or direct OTLP intake.
Reads common GenAI attribute namesContrastYes. GenAI, OpenInference, Langfuse, LangSmith and more.DatadogPartly. GenAI 1.37+ and OpenInference, with conversion on.

Pricing and agents

No per-host or per-span chargeContrastYes. Data scanned, storage and AI credits.DatadogNo. Per host, or per LLM span.
Billed by data scannedContrastYes. $10 per TB after 50 GB free.DatadogPartly. Ingested spans are billed by the GB.
Hosted MCP serverContrastYes. TraceDB tools in the Contrast MCP server.DatadogYes. Datadog MCP Server.
Agents run SQL on tracesContrastYes. MCP tools run SQL on TraceDB.DatadogPartly. DDSQL toolset, in preview.

Running alongside Datadog

  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. Add an otlphttp exporter for Contrast to the OpenTelemetry Collector you already run, next to your Datadog exporter, so both get the same spans.
  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, the median of 5 races. 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. Datadog was not part of this test. The cost figure is the one the Contrast home page states for TraceDB against Snowflake. Datadog details are from its public documentation, checked November 2026: Trace retention, APM billing, Agent Observability, Agent Observability data governance, OpenTelemetry instrumentation, OTLP intake endpoint, DDSQL reference, DDSQL LLM Observability dataset. Product names are trademarks of their owners.