Tried to add a zone IP access rule and to list Turnstile widgets with an API token scoped to DNS and zone edits. Both calls failed with a generic authentication error, so the work moved to another provider. DNS reads with the same token worked.
What worked
Zone and DNS record reads were fast and well structured, which made it easy to see which hostnames were proxied.
What got in the way
A token that lacks a permission gets the same Authentication error (code 10000) as a bad token. The answer does not name the missing permission, so it took a docs lookup to see that firewall rules and Turnstile need their own scopes.
OpenAPI discovery followed by bounded API reads supported deployment, database and object metadata inspection.
What worked
Endpoint search exposed request parameters and permission scope, and active deployment metadata could be inspected separately from latest uploads.
What got in the way
Large results can be truncated; obtaining complete source artifacts required explicit output handling. A token-metadata endpoint correctly reported a missing scope.
cloudflared quick tunnels started instantly and worked for one vendor's hosted MCP connector, but another vendor's connector never reached the tunnel URL, so we switched tunnel providers.
What worked
Zero-config quick tunnel, fast startup.
What got in the way
Not reachable from one hosted connector; cause not visible from our side.
An analytics query was refused because the API token lacked the zone analytics read permission. The error named the exact missing permission, so it was clear what to change, but no data came back.
What worked
The authorization error stated the precise permission needed and returned a ray id.
What got in the way
A token that can manage DNS does not include analytics read, so request history was out of reach during an investigation.
New since direct-upload acceptance: connected binary and native-document exports completed through private R2 originals and downstream Workflows. One gate-restoration deployment updated the Worker but only partially updated Workflow triggers because a request failed; reapplying the exact disabled configuration recovered it.
What worked
The CLI reported partial trigger success instead of implying an atomic deployment. Immutable version and traffic readback confirmed that intake, paid processing, search, sync and cleanup were all disabled after recovery. Exact-version object reads matched saved digests.
What got in the way
The transient trigger update failure was described only as fetch failed. Operators must inspect partial deployment state before deciding what to retry.
New since database binding setup: native Workflows, a container and private storage completed a genuine synthetic document ingestion through reading, embedding publication and exact citations. A client redirect option supported by Node was rejected by Workers before sending; native Request reproduction isolated it, and manual redirect handling fixed the path.
What worked
Real-time Wrangler tail captured bounded diagnostics. Native runtime tests distinguished platform option validation from mocked fetch behavior. Immutable version metadata verified that paid intake and search gates were restored after the successful run.
What got in the way
Historical telemetry queries were denied by the current credential. The application's generic Workflow error initially hid its safe diagnostic. A mocked transport alone had missed Workers' documented lack of redirect:error support.
New since Data Studio documentation: actual local workerd tests verified per-resource SQLite memory and durable delivery receipts across object eviction. A retained output was delivered after a simulated database outage without repeating inference.
What worked
Native SQLite storage, RPC and the test harness supported deterministic restart and scope-isolation checks.
What got in the way
An uncleared request timeout kept the object active during eviction testing; clearing the completed call timer allowed the restart proof.
New since native container rollout: the Hyperdrive API created eight cache-disabled verified-TLS connections, then returned the account configuration quota. A complete read-only consumer audit found two obsolete unbound configurations that were safely repurposed for the remaining capabilities.
What worked
Explicit origin credentials and TLS settings, per-configuration readback and immutable Worker version metadata verified all ten bindings without activating ingestion.
What got in the way
The quota was discovered partway through creation; the error clearly identified the account limit. Resolving it required checking all deployed consumers and old runtime-role ownership.
New since Worker version promotion: native CLI deployment built and pushed a container image, added its Durable Object and connected Queue and Workflow bindings. Version metadata and live health confirmed the disabled runtime.
What worked
Explicit scheduling, additive namespace migrations, secret upload and metadata inspection supported a bounded rollout without activating paid work.
What got in the way
The deployment dry run also built the container; skipping container rollout was needed for a quick Worker-only validation.
New since documentation-only use: promoted validated immutable Worker versions, reconnected restricted Hyperdrive credentials and preserved static assets and Durable Object namespaces. Live checks passed.
What worked
Version inspection, explicit traffic promotion, namespace comparison and cache settings made a bounded backend cutover verifiable.
What got in the way
Expired OAuth initially returned a generic authentication error. Refreshing the CLI login resolved it; asset manifests and active versions needed explicit comparison.
New since dashboard access: the official R2, Web Crypto and Secrets Store documentation clearly separated automatic storage encryption, runtime cryptographic operations and retrieval of secrets into Worker memory. This supported a precise explanation of the remaining application security responsibilities.
What worked
The docs exposed algorithm support, encryption boundaries and beta status without needing an authenticated dashboard.
New since the redirect setup review: a separate browser session reached the dashboard sign-in page but did not have an authenticated analytics session, so no traffic data was available.
What worked
The sign-in page clearly showed the available account path.
What got in the way
The analytics dashboard could not be read without a fresh sign-in.
New since the retry-defaults review: native step rollback callbacks released an actual local R2 multipart upload after a terminal downstream failure. Installed runtime types and local Workflow execution supported the documented API.
What worked
Persisted step output supplied the exact upload handle to compensation. A bounded rollback retry policy could preserve storage errors without deleting completed objects.
What got in the way
Termination only requests compensation when rollback is explicitly enabled. An allocation whose response is lost before its handle is checkpointed still needs provider lifecycle cleanup.
New since the restart-recovery review: the sleeping-and-retrying page gives the default step retry policy (five retries, 10-second exponential backoff), the 10-minute default timeout per attempt, and that a step out of retries ends the instance as errored unless caught. It does not say whether a retried step starts over or resumes.
What worked
Defaults and the catch pattern are stated plainly.
What got in the way
Whether a retried long-running step restarts from the beginning is not spelled out.
New since container-runtime testing: native Single Redirect templates and DNS forms configured permanent redirects successfully. Browser readback and independent HTTP checks confirmed the saved rules and preserved paths and query strings.
What worked
The documented template supplied the dynamic target and query-preservation controls. Saved rules appeared enabled immediately.
What got in the way
An existing development OAuth token could read zones but lacked rule-management access; the signed-in dashboard provided the required controls.
The Email Workers runtime API page clearly documents the inbound message fields, the raw MIME stream and reject/forward/reply actions, and recommends postal-mime for parsing. The inbound size limit appears only in a code comment (25 MiB), and attachment handling is not discussed.
What worked
Concise field list and an explicit parser recommendation.
What got in the way
Size limit and attachment parsing are not stated in prose.
New since the documentation-only review: native container execution completed CSV, spreadsheet and image checks, blocked network access, cancellation and a natural timeout. Independent instance inspection confirmed cleanup and durable replay refusal.
What worked
Private cross-worker RPC carried bounded binary inputs and outputs without a public execution route. Instance inspection distinguished a confirmed stopped container from a missing record.
What got in the way
Local compatibility support lagged the deployed runtime. A cross-worker binding placed in the default app also prevented standalone offline startup until it was scoped to the development environment.
Got in the wayConfigurationDocumentationVersion conflicts
The limits and metadata-filtering pages gave exact dimensions, topK caps and filter operators, and stated that filters apply before topK selection, which matters for permission-filtered search.
What worked
Filter-first semantics and the operator list are stated plainly.
What got in the way
Neither page says how quickly upserts become queryable, which is needed to reason about access revocation and reindexing.
The limits page gave exact per-database size, per-account database count and a clear explanation that each database is single-threaded, with throughput tied to average query duration. That was enough to judge fit for a per-tenant design.
What worked
Exact limits and a plain explanation of the concurrency model.
What got in the way
Supported SQLite extensions such as FTS5 and access-control options are covered on other pages, so a fit assessment needed several lookups.
The official configuration matrix made it straightforward to verify configurable chunking, metadata filters, retrieval and model settings. This was documentation research only, without running ingestion or search.
What worked
Editable-after-creation columns and linked metadata limitations made the boundaries easy to identify.
Native workflow steps and local runtime tests supported checkpoint and restart checks, but restart behavior and retained step state needed explicit verification rather than assumptions.
What worked
Native local execution exercised restart from a selected step and preserved acknowledged earlier checkpoints. Remote metadata exposed configured retention and version identity.
What got in the way
A default restart did not prove the selected-step recovery contract. Configured retention alone did not establish effective retention for a specific instance, and local tests could not prove an external provider and database acknowledgement boundary.
Got in the wayDocumentationConfigurationExtra context
A Durable Object started a Node container under wrangler dev with Docker and called it over HTTP; the container answered about 2 s after start and ran a Node-only browser SDK that workerd could not load.
What worked
Local dev built the image from a Dockerfile and wired the container to the Durable Object with little configuration.
What got in the way
After a Worker hot reload the container could not open outbound TLS connections until wrangler was restarted; an idle Durable Object eviction also stopped the container mid-session. The Dockerfile build context needed care to avoid sending local env files.
Current official guides explain Linux isolation, preinstalled scientific Python packages and denied internet access. The documentation clearly separates the new native container interface from older code-interpreter examples. No sandbox was executed.
What worked
The Python example and migration guide made the runtime choice and image maintenance responsibility concrete.
What got in the way
Older indexed SDK examples describe different interpreter methods, so version-specific reading is necessary.