MongoDB Atlas Stream Processing

MongoDB Atlas Streams

Build, operate, and debug Atlas Stream Processing (ASP) pipelines using four MCP tools from the MongoDB MCP Server.

Prerequisites

This skill requires the MongoDB MCP Server connected with:

• Atlas API credentials (apiClientId and apiClientSecret)

The 4 tools: atlas-streams-discover, atlas-streams-build, atlas-streams-manage, atlas-streams-teardown.

All operations require an Atlas project ID. If unknown, call atlas-list-projects first to find your project ID.

If MCP tools are unavailable

If the MongoDB MCP Server is not connected or the streams tools are missing, see references/mcp-troubleshooting.md (opens in a new tab) for diagnostic steps and fallback options.

Tool Selection Matrix

atlas-streams-discover — ALL read operations

Pagination (all list actions): limit (1-100, default 20), pageNum (default 1). Response format: responseFormat — "concise" (default for list actions) or "detailed" (default for inspect/diagnose).

atlas-streams-build — ALL create operations

Field mapping — only fill fields for the selected resource type:

• resource = “workspace”: Fill: projectId, workspaceName, cloudProvider, region, tier, includeSampleData. Leave empty: all connection and processor fields.
• resource = “connection”: Fill: projectId, workspaceName, connectionName, connectionType, connectionConfig. Leave empty: all workspace and processor fields. (See references/connection-configs.md (opens in a new tab) for type-specific schemas.)
• resource = “processor”: Fill: projectId, workspaceName, processorName, pipeline, dlq (recommended), autoStart (optional). Leave empty: all workspace and connection fields. (See references/pipeline-patterns.md (opens in a new tab) for pipeline examples.)
• resource = “privatelink”: Fill: projectId, privateLinkConfig. Note: PrivateLink is project-level, not workspace-level. workspaceName is not required — omit it. Leave empty: all connection and processor fields.

atlas-streams-manage — ALL update/state operations

Field mapping — always fill projectId, workspaceName, then by action:

• "start-processor" → resourceName. Optional: tier, resumeFromCheckpoint, startAtOperationTime (ISO 8601 timestamp to resume from a specific point)
• "stop-processor" → resourceName
• "modify-processor" → resourceName. At least one of: pipeline, dlq, newName
• "update-workspace" → newRegion or newTier
• "update-connection" → resourceName, connectionConfig. Exception: networking config (e.g., PrivateLink) cannot be modified after creation — delete and recreate.
• "accept-peering" → peeringId, requesterAccountId, requesterVpcId
• "reject-peering" → peeringId

State pre-checks:

• start-processor → errors if processor is already STARTED
• stop-processor → no-ops if already STOPPED or CREATED (not an error)
• modify-processor → errors if processor is STARTED (must stop first)

Processor states: CREATED → STARTED (via start) → STOPPED (via stop). Can also enter FAILED on runtime errors. Modify requires STOPPED or CREATED state.

Teardown safety checks:

• Processor deletion → auto-stops before deleting (no need to stop manually first)
• Connection deletion → blocks if any running processor references it. Stop/delete referencing processors first.
• Workspace deletion → See detailed workflow below (lines 108-111).

atlas-streams-teardown — ALL delete operations

Field mapping — always fill projectId, resource, then:

• resource: "workspace" → workspaceName
• resource: "connection" or "processor" → workspaceName, resourceName
• resource: "privatelink" or "peering" → resourceName (the ID). These are project-level resources, not tied to a specific workspace.

Before deleting a workspace, inspect it first:

1. atlas-streams-discover → inspect-workspace — get connection/processor counts
2. Present to user: “Workspace X contains N connections and M processors. Deleting permanently removes all. Proceed?”
3. Wait for confirmation before calling atlas-streams-teardown

CRITICAL: Validate Before Creating Processors

You MUST call search-knowledge before composing any processor pipeline. This is not optional.

• Field validation: Query with the sink/source type, e.g. “Atlas Stream Processing $emit S3 fields” or “Atlas Stream Processing Kafka $source configuration”. This catches errors like prefix vs path for S3 $emit.
• Pattern examples: Query with dataSources: [{"name": "devcenter"}] for working pipelines, e.g. “Atlas Stream Processing tumbling window example”.

Also fetch examples from the official ASP examples repo when building non-trivial processors: https://github.com/mongodb/ASP_example (opens in a new tab) (quickstarts, example processors, Terraform examples). Start with example_processors/README.md for the full pattern catalog.

Key quickstarts:

Pipeline Rules & Warnings

Invalid constructs — these are NOT valid in streaming pipelines:

• $$NOW, $$ROOT, $$CURRENT — NOT available in stream processing. NEVER use these. Use the document’s own timestamp field or _stream_meta metadata for event time instead of $$NOW.
• HTTPS connections as $source — HTTPS is for $https enrichment or sink only, NOT as a data source
• Kafka $source without topic — topic field is required
• Pipelines without a sink — terminal stage ($merge, $emit, $https, or $externalFunction async) required for deployed processors (sinkless only works via sp.process())
• Lambda as $emit target — Lambda uses $externalFunction (mid-pipeline enrichment), not $emit
• $validate with validationAction: "error" — crashes processor; use "dlq" instead

Required fields by stage:

• $source (change stream): include fullDocument: "updateLookup" to get the full document content
• $source (Kinesis): use stream (NOT streamName or topic)
• $emit (Kinesis): MUST include partitionKey
• $emit (S3): use path (NOT prefix)
• $https: must include connectionName, path, method, as, onError: "dlq"
• $externalFunction: must include connectionName, functionName, execution, as, onError: "dlq"
• $validate: must include validator with $jsonSchema and validationAction: "dlq"
• $lookup: include parallelism setting (e.g., parallelism: 2) for concurrent I/O
• AWS connections (S3, Kinesis, Lambda): IAM role ARN must be registered via Atlas Cloud Provider Access first. Always confirm this with user. See references/connection-configs.md (opens in a new tab) for details.

See references/pipeline-patterns.md (opens in a new tab) for stage field examples with JSON syntax.

SchemaRegistry connection: connectionType must be "SchemaRegistry" (not "Kafka"). Schema type values are case-sensitive (use lowercase avro, not AVRO). See references/connection-configs.md (opens in a new tab) for required fields and auth types.

MCP Tool Behaviors

Elicitation: When creating connections, the build tool auto-collects missing sensitive fields (passwords, bootstrap servers) via MCP elicitation. Do NOT ask the user for these — let the tool collect them.

Auto-normalization:

• bootstrapServers array → auto-converted to comma-separated string
• schemaRegistryUrls string → auto-wrapped in array
• dbRoleToExecute → defaults to {role: "readWriteAnyDatabase", type: "BUILT_IN"} for Cluster connections

Workspace creation: includeSampleData defaults to true, which auto-creates the sample_stream_solar connection.

Region naming: The region field uses Atlas-specific names that differ by cloud provider. Using the wrong format returns a cryptic dataProcessRegion error.

See references/connection-configs.md (opens in a new tab) for the full region mapping table. If unsure, inspect an existing workspace with atlas-streams-discover → inspect-workspace and check dataProcessRegion.region.

Connection Capabilities — Source/Sink Reference

Know what each connection type can do before creating pipelines:

Common connection usage mistakes to avoid:

• ❌ Using $externalFunction as sink with execution: "sync" → Must use execution: "async" for sink stage
• ❌ Forgetting change streams exist → Atlas Cluster is a powerful source, not just a sink
• ❌ Using $merge with Kafka → Use $emit for Kafka sinks

See references/connection-configs.md (opens in a new tab) for detailed connection configuration schemas by type.

Core Workflows

Setup from scratch

1. atlas-streams-discover → list-workspaces (check existing)
2. atlas-streams-build → resource: "workspace" (region near data, SP10 for dev)
3. atlas-streams-build → resource: "connection" (for each source/sink/enrichment)
4. Validate connections: atlas-streams-discover → list-connections + inspect-connection for each — verify names match targets, present summary to user
5. Call search-knowledge to validate field names. Fetch relevant examples from https://github.com/mongodb/ASP_example (opens in a new tab)
6. atlas-streams-build → resource: "processor" (with DLQ configured)
7. atlas-streams-manage → start-processor (warn about billing)

Workflow Patterns

Incremental pipeline development (recommended): See references/development-workflow.md (opens in a new tab) for the full 5-phase lifecycle.

1. Start with basic $source → $merge pipeline (validate connectivity)
2. Add $match stages (validate filtering)
3. Add $addFields / $project transforms (validate reshaping)
4. Add windowing or enrichment (validate aggregation logic)
5. Add error handling / DLQ configuration

Modify a processor pipeline:

1. atlas-streams-manage → action: "stop-processor" — processor MUST be stopped first
2. atlas-streams-manage → action: "modify-processor" — provide new pipeline
3. atlas-streams-manage → action: "start-processor" — restart

Debug a failing processor:

1. atlas-streams-discover → diagnose-processor — one-shot health report. Always call this first.
2. Commit to a specific root cause. Match symptoms to diagnostic patterns:
   • Error 419 + “no partitions found” → Kafka topic doesn’t exist or is misspelled
   • State: FAILED + multiple restarts → connection-level error (bypasses DLQ), check connection config
   • State: STARTED + zero output + windowed pipeline → likely idle Kafka partitions blocking window closure; add partitionIdleTimeout to Kafka $source (e.g., {"size": 30, "unit": "second"})
   • State: STARTED + zero output + non-windowed → check if source has data; inspect Kafka offset lag
   • High memoryUsageBytes approaching tier limit → OOM risk; recommend higher tier
   • DLQ count increasing → per-document errors; use MongoDB find on DLQ collection See references/output-diagnostics.md (opens in a new tab) for the full pattern table.
3. Classify processor type before interpreting output volume (alert vs transformation vs filter).
4. Provide concrete, ordered fix steps specific to the diagnosed root cause. Do NOT present a list of hypothetical scenarios.
5. If detailed logs are needed, direct the user to the Atlas UI: Atlas → Stream Processing → Workspace → Processor → Logs tab.

Chained processors (multi-sink pattern)

CRITICAL: A single pipeline can only have ONE terminal sink ($merge or $emit). When users request multiple output destinations (e.g., “write to Atlas AND emit to Kafka”), you MUST acknowledge the single-sink constraint and propose chained processors using an intermediate destination. See references/pipeline-patterns.md (opens in a new tab) for the full pattern with examples.

Pre-Deploy & Post-Deploy Checklists

See references/development-workflow.md (opens in a new tab) for the complete pre-deploy quality checklist (connection validation, pipeline validation) and post-deploy verification workflow.

Tier Sizing & Performance

See references/sizing-and-parallelism.md (opens in a new tab) for tier specifications, parallelism formulas, complexity scoring, and performance optimization strategies.

Troubleshooting

See references/development-workflow.md (opens in a new tab) for the complete troubleshooting table covering processor failures, API errors, configuration issues, and performance problems.

Billing & Cost

Atlas Stream Processing has no free tier. All deployed processors incur continuous charges while running.

• Charges are per-hour, calculated per-second, only while the processor is running
• stop-processor stops billing; stopped processors retain state for 45 days at no charge
• For prototyping without billing: Use sp.process() in mongosh — runs pipelines ephemerally without deploying a processor
• See references/sizing-and-parallelism.md for tier pricing and cost optimization strategies

Safety Rules

• atlas-streams-teardown and atlas-streams-manage require user confirmation — do not bypass
• BEFORE calling atlas-streams-teardown for a workspace, you MUST first inspect the workspace with atlas-streams-discover to count connections and processors, then present this information to the user before requesting confirmation
• BEFORE creating any processor, you MUST validate all connections per the “Pre-Deployment Validation” section in references/development-workflow.md (opens in a new tab)
• Deleting a workspace removes ALL connections and processors permanently
• After stopping a processor, state is preserved 45 days — then checkpoints are discarded
• resumeFromCheckpoint: false drops all window state — warn user first
• Moving processors between workspaces is not supported (must recreate)
• Dry-run / simulation is not supported — explain what you would do and ask for confirmation
• Always warn users about billing before starting processors
• Store API authentication credentials in connection settings, never hardcode in processor pipelines

Reference Files