npx skills add ...
npx skills add getsentry/sentry --skill cell-architecture
npx skills add getsentry/sentry --skill cell-architecture
Reference and active migration guide for Sentry's cell architecture. Explains what cells and localities are and why they're different, how requests reach cells via Synapse API routing, ingestion routing, and the control silo gateway, and how to safely query cross-cell data without silently missing results. The migration section covers how to do migration work: draining the URL_NAME_TO_ACTION registry in test_urls.py to zero (with a recipe for each action type), rolling deploy safety and the two-phase pattern required by independent sentry/getsentry deploys, and the region -> cell rename including what not to rename (DB columns, AWS refs, uptime regions, billing address). Also documents known issues with proposed fixes: integration TeamLinkageView routing, Jira cross-cell fan-out, and relocation endpoint routing.
Status: Active migration in progress. Migration-specific sections should be removed once complete, leaving a stable architecture reference.
These are two different layers of the architecture.
Cell — a self-contained Sentry deployment that owns a subset of organizations. Each cell
runs its own full stack — Getsentry, Snuba, Seer, Relay, Kafka, Symbolicator, and others —
on an isolated network with no direct cell-to-cell communication. OrganizationMapping.cell_name
records which cell an org lives in. See Paths Into a Cell for how cells
communicate with the outside world.
Locality — a named collection of cells, representing either a data residency zone (for
multi-tenant customers, e.g. "us", "de") or a dedicated deployment for a single customer
(e.g. s4s2). Multi-tenant customers choose a locality when creating an organization;
single-tenant localities are provisioned privately and not customer-selectable. Each locality
maps to a subdomain (us.sentry.io, de.sentry.io or s4s2.sentry.io).
Note: "region" is the old name for "cell". The codebase is actively being migrated. See Active Migration for details.
There are three high-level paths by which requests or data reach a cell.
{locality}.sentry.ioSynapse (getsentry/synapse) routes each request to
the correct getsentry cell within a locality, using an org to cell mapping it caches from control's
OrganizationMapping.
For Synapse to route a request, the URL must contain organization_id_or_slug. The canonical
shape is /api/0/organizations/<organization_id_or_slug>/... — getsentry/tests/getsentry/test_urls.py
enforces this and fails CI for any @cell_silo_endpoint that doesn't conform.
When generating URLs for org-scoped resources, use org.locality.to_url(path) — derives the
correct locality URL across SaaS, self-hosted, single-tenant, and dev deployments:
For single-cell localities Synapse is an optional pass-through.
ingest.{locality}.sentry.ioA separate Synapse deployment routes each ingestion request to the correct cell using a public key to cell mapping it caches from control. The project config returned by Synapse includes the cell's Relay URL directly, which high-volume Relay deployments use to bypass Synapse for subsequent high-volume submission.
DSN hosts vary — legacy formats carry no locality or org info — so the ingest-router always routes by public key, which is present in every request regardless of DSN format.
The o{org_id}.ingest.{locality} subdomain also serves a second routing purpose: the user feedback embed widget (sentry-error-page-embed) posts to sentry.io with the project's DSN as a query param. Since the URL has no org slug, the API gateway can't route it the normal way — instead it parses the locality out of the DSN host. If the DSN was issued in a legacy format without a locality, the gateway can't determine the cell and the embed widget breaks silently for that org.
For single-cell localities the ingest-router is an optional pass-through.
sentry.iosentry.io is the getsentry control silo deployment. Control communicates with cells through three mechanisms:
API Gateway (ApiGatewayMiddleware -> apigateway.py) — synchronously proxies
org-scoped API requests that arrive at sentry.io but belong on a cell:
get_cell_for_organization()sentry-error-page-embed) -> parse locality from DSN subdomainREGION_PINNED_URL_NAMES proxy to the monolith cell (the original US cell). Legacy endpoints like /api/0/issues/{id}/ have no org slug so the gateway can't resolve the cell dynamically; they always route to US and will 404 for issues in other cellsIntegration webhook forwarding (IntegrationControlMiddleware -> BaseRequestParser) —
inbound webhooks arrive at control, which identifies target cells via OrganizationIntegration. organization_mapping_service, then delivers using one of three strategies:
WebhookPayload async queue (default; GitHub, most providers) — creates one record per
target cell, returns 202 immediately; background worker delivers with retriesRPC — synchronous cross-silo calls. Each service is local to one silo; the other calls it remotely. For creating or modifying RPC services -> hybrid-cloud-rpc skill.
Dev / Monolith mode: In local development, all data lives in a single DB and cross-cell bugs are invisible locally — they only surface in a siloed staging or production environment.
A user can belong to organizations in different cells. Any query that filters org-scoped data through a cross-cell relationship — membership, team assignment, or similar — silently returns incomplete results when run inside a single cell: it only sees the orgs that live in that cell.
OrganizationMemberMapping (control) — (user_id, organization_id, role, invite_status)
for every member across all cells. The canonical user->org index; start here.OrganizationMapping (control) — maps organization_id -> cell_nameorganization_service.get_organization_by_id() — fetches org details; routes to the
correct cell automaticallyfind_cells_for_user(user_id) — returns cell names containing the user's orgs (use when
you need to fan out to cells but don't need membership data; calls OrganizationMemberMapping
internally)| Pattern | When to use |
|---|---|
| Control index + RPC fan-out | Real-time accuracy required; moderate org count per user. Use OrganizationMemberMapping to get org IDs, then organization_service RPC per org (or batch by cell via OrganizationMapping). |
| Denormalize to control silo | High QPS, paginated lists, eventual consistency acceptable. Replicate needed fields to a @control_silo_model via outboxes — single local query, no RPC. See hybrid-cloud-outboxes skill. |
| API gateway fan-out | Temporary stopgap only — breaks pagination, latency = slowest cell. |
Only cell-scoped API XHRs (/api/0/organizations/{slug}/*) cross to Synapse on :13000; UI HTML and control API stay on the sentry devserver at :8000. The env var gates a block in src/sentry/conf/server.py that sets system.region-api-url-template to the Synapse port and flips system:multi-region. The frontend reads regionUrl from the bootstrap blob and routes cell XHRs there per static/app/api.tsx:715.
bin/send-cell-test-event.py exercises the ingestion path: it resolves a DSN from the
local dev DB and sends an envelope through the edge relay (:7899), which forwards to
relay-cell -> Kafka -> Sentry.
Watch it route with docker logs -f sentry-relay-edge-1.
Remove this section once the migration is complete.
Use TODO(cells) to track deferred migration work so it's searchable:
Public @cell_silo_endpoint URLs must include organization_id_or_slug so Synapse can route them — see Paths Into a Cell. getsentry/tests/getsentry/test_urls.py enforces this in CI and fails if a non-conforming URL has no registered plan.
The goal is to drain URL_NAME_TO_ACTION to zero. Each entry is a non-conforming endpoint with a label describing what needs to happen to it. When you fix one, remove it from the map. No new non-conforming entries should be added.
Endpoints decorated with @internal_cell_silo_endpoint are exempt from the check.
Actions and what fixing them looks like:
TO_BE_REPLACED_WITH_ORG_SCOPED_VARIANT — add organization_id_or_slug to the URL path so Synapse can route itTO_BE_REPLACED_WITH_CELL_SCOPED_VARIANT — _admin staff-only endpoint; scope it to a cell by adding a cell ID parameter (pattern: ^api/0/_admin/cells/(?P<cell_id>[^/]+)/)TO_BE_CONTROL_ONLY — move the endpoint to @control_silo_endpointTO_BE_INTERNAL_ONLY — change to @internal_cell_silo_endpoint; this exempts it from the routing requirementTO_BE_STANDARDIZED — fix a URL typo or convention mismatch (e.g. missing org slug support, path not following existing conventions)TO_BE_SELF_HOSTED_ONLY — move behind a self-hosted guard so it's not reachable in SaaS cellsTO_BE_DEPRECATED — endpoint is being removed. Before touching anything: search the getsentry GitHub org for the URL name and path to find all callers. Then add a X-Sentry-Deprecation-Date response header while still live (see _add_deprecation_headers in sentry/api/helpers/deprecation.py). When fully removed: delete from src/sentry/api/urls.py, the view file, URL_NAME_TO_ACTION, and src/sentry/apidocs/ if public.TO_BE_DELETED - not a public endpoint, can be removed immediately without deprecationTO_BE_BROKEN — known broken, no owner; deferred indefinitely (only relocation, do not add entries)TO_BE_INVESTIGATED_ECOSYSTEM_TEAM — deferred until the ecosystem team determines the right fixNever assume atomic deployment. There are two independent deployment boundaries to consider:
Sentry rolling deploy — old and new sentry pods run simultaneously during rollover. Any change to a wire format, RPC method signature, DB schema, or serialized field must be safe for old pods to receive from new pods, and vice versa.
Sentry <-> getsentry — sentry and getsentry are always deployed independently and at different times. Anything sentry exports that getsentry imports (classes, functions, constants, decorators) must remain importable under its old name until getsentry has deployed the update. Keep the old name as an alias; never do a hard rename in a single deploy.
The required pattern for any breaking change is two phases:
Phase 1 — deploy backward-compatible code: add the new name/field/format alongside the old. Both old and new versions of each service can handle both.
Phase 2 — deploy cleanup: remove the old name/field/format once both sentry and getsentry are on the new code.
This applies to:
db_column="old_name" to avoid a schema migration entirely — the DB column stays unchanged and is safe across rolling deploysname= string is serialized into Kafka/the broker; in-flight tasks carry the old name and old kwarg names. Keep the old name registered during the transition (via alias= on @instrumented_task, or by keeping the old task as a shim that calls the new function directly). Remove it in a follow-up deploy once the queue has drained.The codebase is actively migrating from "region" to "cell" terminology.
New code must not use "region" to mean "cell" or "locality" — this applies to all Python symbols, variable names, function names, class names, comments, log keys, and docstrings.
Before renaming, identify which concept the code is referring to:
cell_name, get_cell_for_organization, CellSiloClient, etc.)"us", "de") -> locality
(locality, org.locality, Locality, etc.)Most existing region usage maps 1:1 to cell, but some (especially where the code handles
"us" / "de" subdomain routing or customer-facing zone selection) maps to locality. Check
the context before renaming.
Do NOT rename:
db_column="region_name" — DB schema stays for backward compatibility; only the Python attribute name changesaws-lambda.host-region, etc.), such as in AWS integrationsregion in getsentry — postal address state/province for tax, etc, unrelated to cell infrastructureRename with caution — sentry and getsentry deploy independently on different schedules, so both old and new names must work simultaneously during the transition:
metrics.incr("region.*")) — dashboards and alerts reference metric names; emit both old and new names until getsentry dashboards are updated, then drop the old.options.get("hybridcloud.regionsiloclient.*")) — values are set in production via sentry-options-automator in getsentry; register the new key alongside the old, migrate getsentry config, then remove the old key in a later deploySpecific broken areas with a description of the problem and the proposed fix. Remove each entry once resolved.
TeamLinkageView (link-team, unlink-team) is broken in two ways: it uses signed URLs with integration_id instead of org slug so Synapse can't route it, and it queries OrganizationMember.get_for_integration() (cell-silo) so it only sees orgs in the current locality anyway.
The fix:
@control_silo_view — makes the URL reachable without SynapseOrganizationMember.get_for_integration() with OrganizationMemberMapping + integration_service.get_organization_integrations() for cross-cell membershipTeam and write ExternalActor via RPC into the correct cell — requires new RPC methods (see hybrid-cloud-rpc skill)IdentityLinkageView is already @control_silo_view and doesn't need migration.
The Jira sidebar panel ("Sentry -> Linked Issues") is blocked from multi-cell installation via is_cell_restricted = True instead of being fixed properly. The descriptor points to JiraSentryIssueDetailsView (@cell_silo_view), which queries a single cell's DB. JiraSentryIssueDetailsControlView (@control_silo_view) already exists at /extensions/jira/issue-details/{issue.key}/ with correct find_cells_for_orgs fan-out — the fix is to point the descriptor at that URL and remove is_cell_restricted.
All relocation endpoints are @cell_silo_endpoint with no org slug — /relocations/, /publickeys/relocations/, and the UUID-based management endpoints (/relocations/{uuid}/abort/ etc.) — so Synapse can't route any of them in a multi-cell locality. Likely fix: UUID-based endpoints via a control-silo uuid -> cell lookup; list and public key endpoints follow the org-creation pattern (locality selection -> control picks a cell). Will be addressed as part of the Monarch project (org migration between cells).
# Looks correct, silently wrong in multi-cell — misses orgs in other cells:
Organization.objects.get_for_user_ids({user_id})
Project.objects.filter(teams__organizationmember__user_id=user_id)devservices up --mode cell-routing # brings up synapse + deps
SENTRY_CELL_ROUTING=1 sentry devserver --client-hostname=dev.getsentry.net # or devservices servebin/send-cell-test-event.py # internal project (id 1)
PROJECT_ID=42 bin/send-cell-test-event.py # different project
TARGET=127.0.0.1:7900 bin/send-cell-test-event.py # straight to relay-cell, bypass edge# TODO(cells): rename metric to "cell.foo" once getsentry dashboards are updated
metrics.incr("region.foo", tags=metric_tags)