Context Graph

The context graph is read through the same public SigV4 API as the rest of Omni: service cloudwatch-omni, signing name cloudwatch, invoked as aws cloudwatchomni get-context-graph from the AWS CLI (through the aws___call_aws tool or a shell). See programmatic-access.md for endpoints, SDK access, and what to do when the CLI reports the service as unsupported.

Contents

• Dependency / blast-radius question — facts you MUST surface
• What the context graph is
• GetContextGraph API
• Reading the graph
• Using the graph in an investigation
• Gotchas

Dependency / blast-radius question — facts you MUST surface

“What does <service> depend on and which is broken,” “what’s broken downstream,” “what is the blast radius,” “who is affected if <service> fails,” “<service> is slow or erroring — which way do I walk the graph,” and “what does CALLS / ACCESSES / RUNS_ON mean” are observability-data questions answered from the context graph, not from the resource’s control plane. Answering with the methodology is an output contract — surface every item below that applies, whether or not you also pull live numbers and even if no matching node exists yet. For a direction or edge-semantics question the deliverable IS the explanation below, stated in the answer text, whether or not GetContextGraph is run:

• Edges are directed from → to, and from depends on to. Downstream of X = the edges whose from is X (X’s own edges[]); upstream of X = the edges whose to is X (edgeFilters.to = <X's nodeId>, because node.edges[] is outbound only).
• A slow or erroring service is a DOWNSTREAM question. The cause of X’s latency or errors lies among X’s dependencies, so walk downstream, one hop at a time. Upstream tells you blast radius — who is affected — not cause.
• A blast-radius / who-is-affected question is the REVERSE — traverse UPSTREAM (everything that transitively depends on <service>).
• Say what each edge type means and when to follow it: CALLS — from invokes to, a service dependency; follow it when tracing request-path latency or errors downstream. ACCESSES — from reads or writes datastore to (database, queue, cache); follow it when a service is slow or erroring on I/O. RUNS_ON — from executes on host or compute to; follow it when suspecting an infrastructure cause shared by co-located services.
• Raise depth one hop at a time rather than asking for depth: 3 up front, and query telemetry at each hop before expanding — the graph localizes, queries confirm.
• Compare each edge’s lastObservedAt with the incident window. An edge is returned if it was observed anywhere in the query window; one last seen well before the incident is history, not a live call path.
• A to nodeId that is not on the current page is not a missing node. Pagination is by node and edges are nested under their from node, so collect every page (and check depth) before treating a dangling to as absent.
• An edge with no error fields is unreported, not proven healthy. Only conclude “healthy” from telemetry, never from a missing errorCode / httpStatusCode.
• Confirm the candidate path with traces — parent-to-child spans mirror CALLS edges, so query for slow or erroring spans along exactly that path to establish which hop originates the latency or error.
• Use GetContextGraph, never X-Ray GetServiceGraph. The Omni topology operation is GetContextGraph; if the CLI/SDK errors with an unknown cloudwatch-omni service, that is a client-version issue (upgrade botocore/AWS CLI), not a reason to switch products.

The detailed traversal procedure, direction definitions, and symptom-vs-cause ranking are in Reading the graph and Using the graph in an investigation below.

What the context graph is

The context graph is a directed graph of nodes (the things that run) and edges (the observed relationships between them), derived from telemetry and other discovery sources flowing into a Space. It is the answer to “what talks to what” for one account in one Region, over a time window you choose.

Nodes

Every node has a nodeType, one of:

Orthogonal to nodeType, a node may carry a coarser nodeProperties.category: GEN_AI_AGENT, GEN_AI_MODEL, DATABASE, MESSAGING_QUEUE, COMPUTE, STORAGE, NETWORK. This is how AI agents and the models they call show up in the same graph as ordinary services. The model documents category as absent on most nodes today, so treat its absence as “not reported”, not “not a database”.

A node’s identity lives in nodeProperties — region, cloudProvider (defaults to "aws"), sourceAccountId, namespace, plus category and stage. The first four are the node’s merge key: a service seen from several sources (traces, flow logs, CloudTrail) is resolved into one node, its resolved name wins, and the names it was merged away from appear in alternateNames.

Edges

Every edge is directed, from → to (both are nodeIds), with an edgeType:

edgeProperties carries whatever the producing source reported: errorCode, httpStatusCode, httpMethod, protocol, sourcePort/destinationPort, blocked, serviceInitiated, and trafficStats (bytes, packets, flows, sentBytes, receivedBytes). The presence of errorCode means the edge exists and the dependency was OBSERVED failing on that call — a proven failure within the window, not necessarily one happening right now (check the edge’s lastObservedAt before treating it as a live, current failure). The converse does not hold: an absent errorCode is not evidence of health. It means the producing source did not report one, not that the call succeeded, so a dependency cannot be called healthy from a missing field. The same reading applies to the other optional members — blocked absent means the edge did not come from network-flow data, not that traffic was allowed, and nodeProperties.category is absent on most nodes today. Conclude “healthy” only from actual telemetry (the caller’s spans, metrics, or logs), and check the edge’s lastObservedAt before trusting it as a live call path — an edge seen only early in the window may be history, not a live call. Most edges carry only a few of these members.

Where nodes and edges come from

Each node and edge lists the sources that contributed it and the signalTypes observed on it:

• sources — TELEMETRY (spans and metrics emitted by instrumented code), VPC_FLOW_LOG, ELB_ACCESS_LOG, CLOUDFRONT_ACCESS_LOG, S3_ACCESS_LOG, WAF_ACCESS_LOG, CLOUDTRAIL, IAM_POLICY, CONFIG, AWS_INTEGRATION, CODE_SEMANTICS, AZURE_VNET_FLOW_LOG.
• signalTypes — LOGS, METRICS, TRACES, CONFIG, UNKNOWN.

The request-path edges an investigation cares about (CALLS, ACCESSES) come overwhelmingly from traces: a parent span in service A with a child span reaching service B is a CALLS edge A → B, and the span’s status and attributes are what populate errorCode / httpStatusCode. Vended metrics contribute nodes and the per-node metadata.metrics list. Network-derived edges (flow logs, access logs) add protocol, ports, blocked, and trafficStats. An edge that merged across sources reports the earliest firstObservedAt and latest lastObservedAt any source saw.

How it relates to a Space

The graph is what one account’s Space in one Region can see. A GetContextGraph request carries no Space identifier: it is scoped by the credentials you sign with (the caller’s account) and the Region you send it to. Because an account has at most one Space per Region (see concepts.md), the Region of the call is the Space selection. Calling the wrong Region returns a different — usually empty — graph, not an error.

Where a Space centralizes telemetry forwarded from other accounts, those services appear as nodes whose nodeProperties.sourceAccountId differs from the caller’s account. Read account and Region for later queries off the node — never assume they match the credentials you are using.

GetContextGraph API

GetContextGraph queries the context graph with filtering, traversal, and pagination. It is read-only. The caller’s access grant (or IAM policy) must allow cloudwatch:GetContextGraph; the model does not document a resource-level ARN for it, so scope it at the Space rather than inventing a resource ARN.

Request

NodeFilters — criteria for selecting the starting nodes:

EdgeFilters — criteria for the edges returned:

KeyFilter is { key, values[] }. Multiple KeyFilters in one list are AND’ed; multiple values inside one KeyFilter are OR’ed. Each value may be an exact match, a negation (!value), or a wildcard (*value*, value*, *value).

Response

Edges are not a separate top-level list. Each Node carries its outbound edges in node.edges[], every one of them with its own from (equal to that node’s nodeId) and to. Pagination cursors advance over nodes; each page includes the edges connecting the nodes on that page. Treat nodes as the collection you paginate and edges as relationship data attached to them.

Node:

Edge:

metadata (only with includeMetadata: true) is what lets you pivot from a graph element to the exact telemetry behind it:

• metrics[] — each metric observed on the node: name, preferredStat (e.g. "p99", often absent), metricType (gauge, sum, histogram, …), semantics.description / semantics.unit, and attributes — the raw, stored attribute values (service.name, service.namespace, cloud.account.id, cloud.region, …) that select this metric’s series. These are deliberately not the node’s normalized identity, because a merged node can carry different raw values per metric.
• semantics — on nodes only: purpose, language, framework, kind, repository.
• logs[] / traces[] — per-signal query selectors: a list of blocks that are OR’ed together, each an AND of exact store column → raw values. logs is node-level only; traces appears on both nodes and edges.

Errors

ValidationException (bad range, endTime before startTime, out-of-range depth/maxResults), AccessDeniedException (missing cloudwatch:GetContextGraph, or the account is not enabled for Omni Intelligence), ThrottlingException (retryable), InternalServerException.

CLI form

Nested structures are passed as JSON. Timestamps are ISO-8601. Because the operation is modeled as paginated, the CLI auto-paginates by default — pass --no-paginate to get exactly one page, and --max-items to cap the total.

The response from the first call looks like this (abridged from the model’s own example):

Reading the graph

Direction: upstream, downstream, blast radius

Edges point in the direction of dependency: from depends on to. Downstream = follow edges whose from is X; upstream = edges whose to is X. Walk downstream to find a cause; walk upstream to size blast radius.

• Downstream of X — the nodes X’s own edges[] point to. These are X’s dependencies; if one of them is failing, X will look degraded.
• Upstream of X — the nodes whose edges point to X. These are X’s callers; if X is failing, they will look degraded. node.edges[] is outbound only, so callers are found with edgeFilters.to = <X's nodeId> (the third CLI example above), not by reading X.
• Blast radius of X — everything transitively upstream of X: the set of services whose degradation X alone would explain. A node with a large blast radius and a failing edge beneath it is the shape a root cause usually takes.

The model describes depth as hops traversed “out from” the matched nodes, but the traversal is undirected: each hop follows edges in both directions, so depth: 1 on X returns X’s callers as well as its dependencies, and every intermediate node from hop 0 up to depth (max 3) is included, not only the outermost hop. edgeFilters narrow which edges are returned, not which are traversed. To isolate callers, filter with edgeFilters.to = <X's nodeId> rather than sifting the depth output.

Symptom vs. cause

• A symptom is an upstream caller that looks degraded only because something it CALLS or ACCESSES downstream is failing.
• A cause is a shared downstream dependency that all failing paths traverse.

Prefer the cause. Rank candidates, in order, by:

1. Downstream fan-in explained — the deepest node whose failure accounts for the most affected upstream services.
2. Shared dependency across multiple independently failing services.
3. Temporal precedence — the node whose failing edges or anomalies began first. On the graph itself, compare firstObservedAt of the failing edge (or of a newly-appeared node) against the incident window.

Name the deepest node that explains the most downstream symptoms as the primary hypothesis — not the first degraded node you saw — and state the traversal path (node names and edge types) as the supporting evidence.

Hop-by-hop traversal

1. Locate the alarming node — the service named by the alert, the Omni Intelligence insight shown in the console, or the user’s symptom. Fetch it with nodeFilters.name (or nodeId), depth: 1, and a narrow window around the incident.
2. Read the edges. Any edgeProperties.errorCode or a non-2xx httpStatusCode on a CALLS/ACCESSES edge is a live failing dependency — follow it first. Edges with no error fields are not proven healthy; they are unreported.
3. Walk outward along CALLS / ACCESSES — not just immediate neighbors. At each hop, query telemetry for correlated errors or latency at that node before expanding further (see the next section for how).
4. Go deeper only where signal appears. Stop a branch that is clean. Raise depth one hop at a time rather than asking for depth: 3 up front — a dense graph at depth 3 is hundreds of nodes and hides the path you care about.
5. Confirm along the path with traces. Parent → child spans mirror CALLS edges, so once a candidate path exists, query for slow or erroring spans along exactly that path to establish which hop originates the latency or error.

Using the graph in an investigation

The graph localizes; queries confirm. Alternate between them, spending the cheap call (graph) before the expensive one (telemetry scan).

1. Start from the symptom. An alert firing (alerts.md) names a query and a Space; a CloudWatch Omni Intelligence insight or anomaly in the console names the affected service(s), severity, and time; a user names a service. Any of these gives you a service name, a Region, and a time window. Listing insights and anomalies is not part of the public API today — read them in the console, then carry the affected service names into the graph call.

2. Fetch the neighborhood. get-context-graph with nodeFilters.name, depth: 1, includeMetadata: true, and a window of roughly the incident ± 15 minutes. Take nodeProperties.region and nodeProperties.sourceAccountId from the returned node for every subsequent call — do not guess them from the credentials you hold.

3. Pick the next node by the rules above: follow the edge that carries an errorCode/httpStatusCode, else the shared ACCESSES datastore, else the CALLS edge with the most operations.

4. Pivot to telemetry at the suspected node, using the metadata block so your query matches what is actually stored rather than the graph’s normalized name:

   • Traces and logs (SQL) — metadata.traces[] / metadata.logs[] give the exact column → value selectors; alternateNames tells you which raw service.name values to include. Write the query per query/sql-logs-traces.md, always with a `@timestamp` range, and discover the real field names with EXPLAIN (ANALYZE_FIELDS) before filtering on anything but @-fields. For example, to find erroring spans at the suspected node inside the window:

     sqlCopy

     SELECT `@timestamp`, `@record`
     FROM traces.default
     WHERE `@timestamp` BETWEEN to_timestamp_nanos('2026-09-16T00:00:00Z')
                            AND to_timestamp_nanos('2026-09-16T01:00:00Z')
       AND resource['attributes']['service.name'] = 'payments-service'
     ORDER BY `@timestamp` DESC
     LIMIT 50

     Then narrow with the status / error attributes that schema discovery shows are present for this service.

   • Metrics (PromQL) — metadata.metrics[] gives the metric name, its preferredStat, and the raw attributes (service.name, service.namespace, cloud.region, …) to use as label matchers. Write the query per query/promql-metrics.md.

   • An edge’s operations[] (e.g. "POST /charges") is the operation name to filter spans or metrics on when you want just that call path.

5. Decide, then expand or stop. Signal at this node → fetch its neighborhood (nodeFilters.nodeId = <that node>, depth: 1) and repeat from step 3. No signal → this branch is a symptom; go back to the last node with signal and try its next edge.

6. Report the deepest node that explains the most upstream symptoms, the path you walked (checkout-service —CALLS→ payments-service —ACCESSES→ payments-db), the edge evidence (errorCode, httpStatusCode, firstObservedAt), and the query that confirmed it. If the cause is a REMOTE_SERVICE node you cannot query, say so: the evidence is the failing edge and the caller-side spans, and the fix is on the caller’s side (retries, timeouts, fallbacks) or with the remote owner.

Gotchas

• Empty nodes is usually scope, not absence. In order of likelihood: the Region is wrong (a Space is per Region — the call went to a Region with a different or no Space); the time window does not overlap any observation (firstObservedAt/lastObservedAt are minute-granular; widen the window before concluding); nodeFilters.name is an exact match and the stored name differs (drop name, filter on namespace or nodeType, and look at alternateNames); nothing is instrumented, or telemetry is not reaching this Space yet (see the setup order in concepts.md — instrumentation started before a Space exists delivers nowhere you can see).
• AccessDeniedException on an otherwise-correct call means either the principal lacks cloudwatch:GetContextGraph or the account is not enabled for Omni Intelligence. Neither is fixed by changing the request.
• Edges can point at nodes that are not on this page. Edges are nested under their from node and pagination is by node, so a to may reference a nodeId you have not received yet — or one past your depth. Collect every page (or let the CLI auto-paginate) before treating a dangling to as a missing node.
• maxEdgesPerNode silently truncates hubs. A shared datastore or gateway with more outbound edges than the cap loses some of them from the response. If a dependency you expect is missing, raise the cap (max 50) or query it directly with edgeFilters.from/to.
• Stale edges. An edge is returned if it was observed anywhere in your window. Compare its lastObservedAt with the incident: an edge last seen well before the incident is history, not a live call path — and a CALLS edge that stops appearing mid-window can itself be the signal (the caller stopped reaching the dependency).
• Absent is not negative. errorCode absent means the source did not report one, not that the call succeeded. blocked absent means the edge did not come from network-flow data, not that traffic was allowed. category is absent on most nodes. Only conclude “healthy” from telemetry, never from a missing field.
• namespace is a logical service grouping, not a CloudWatch metric namespace. Do not paste it into a metrics query as one. The metric’s own metadata.metrics[].attributes["service.namespace"] is what selects series.
• Merged nodes do not match stored telemetry by name. The graph’s name is the resolved identity; the store still holds the raw values. Use alternateNames, metadata.traces[]/logs[] selectors, and metadata.metrics[].attributes for the values your queries must match.
• telemetryAttributes filters do not narrow results today (both on edgeFilters and, per the model, nodeFilters). Filter server-side on name, namespace, nodeType, sourceAccountId, region, or stage, and post-filter on attributes client-side.
• REMOTE_SERVICE nodes have only a caller’s-eye view. There is no telemetry of their own in this Space to pivot to. Their edges’ errorCode/httpStatusCode and the calling service’s spans are all the evidence you will get.
• Cross-account nodes need the node’s account, not yours. When nodeProperties.sourceAccountId differs from the caller’s account, the telemetry was forwarded in. Queries against this Space still work, but any per-account follow-up (IAM, resource configuration) must target sourceAccountId.
• includeMetadata is not free. It adds a lookup per returned node. Use it on the small depth: 1 neighborhood you are about to pivot from, not on a depth: 3 sweep of the whole Space.