keepalive/plans/reports/review-261009-1025-adapter-consolidation-and-new-services.md

type, topic, created_at, project

type	topic	created_at	project
review-report	adapter-consolidation-and-new-services	2026-10-09T10:25:00+07:00	keepalive

Review Report: Adapter Consolidation and New Services

Summary

This review covers the whole tiennm99/keepalive module: main.go, config.go, runner.go, service_name.go and every adapter under adapter/. It confirmed ten findings. Two are high or high-leaning: the Couchbase increment resets the counter on any read error, and malformed connection URLs leak passwords into logs. Four are medium reliability gaps in the runner and config loading, and four are low.

On consolidation, one merge is clearly worth doing: serve both redis and valkey with github.com/redis/go-redis/v9 and drop github.com/valkey-io/valkey-go. A scratch build showed this cuts the stripped binary from 36,503,712 B to 27,394,208 B (about 25 percent). It also removes a real connect failure against servers without client-side caching. The merge needs a URL normalizer that is more careful than the first design, and the corrections are listed below. Merging postgresql and mysql behind one database/sql dialect struct is possible, but the review recommends skipping it on KISS grounds. Swapping lib/pq for pgx is rejected.

On new services, many hosted free tiers already work with the existing adapters and only need README entries: Redis Cloud, Upstash, YugabyteDB Aeon, CockroachDB Basic, Azure DocumentDB, FerretDB Cloud and Aiven. Among services that need new code, the most valuable additions are a small set of plain net/http adapters (Qdrant, Zilliz, Neo4j Aura, Astra DB, Turso, Weaviate, InfluxDB), an AMQP adapter for CloudAMQP, a Kafka adapter for Aiven Kafka, an MQTT adapter for EMQX Serverless, and a pure-Go Oracle adapter for the Oracle Autonomous Database. Every candidate is listed for the user to decide on, and nothing has been added.

Confirmed review findings

Every finding below was confirmed by reading the code, and most were also reproduced with a scratch probe outside the repo. Severity reflects the verifier's correction where one was made.

#	Severity	File:line	Finding	Fix
1	High	adapter/couchbase.go:110	Increment discards its context (_ context.Context) and treats any Get error as "missing document", so a timeout, temporary failure, RBAC error or missing collection silently resets the counter to 1. Get then Upsert is not atomic. The WAN development profile sets a 20s KV timeout, so one tick can block for about 40s and ignore the runner's 3s tick timeout, and on SIGTERM wg.Wait() can outlast Docker's default 10s stop grace period.	Replace both calls with one atomic a.coll.Binary().Increment(a.docID, &gocb.IncrementOptions{Delta: 1, Initial: 1, Context: ctx}) and return int64(res.Content()). The document that ensureDocument seeds holds the digit text 0, so the first tick still returns 1.
2	Medium (reported high)	runner.go:36	When a URL fails net/url parsing, lib/pq, redis.ParseURL and valkey.ParseURL return a *url.Error that contains the full raw URL, password included. The runner logs it every 10s. A probe reproduced it for all three adapters with a password containing %9z. An unencoded # leaks only the part of the password before the #. Severity is lowered because only a malformed URL triggers it and the logs belong to the credential owner, but the line repeats and reaches Coolify and log-shipper storage.	Validate the url key once at config load. Alternatively, have the URL-based adapters share one helper that uses errors.As to detect a *url.Error and returns only urlErr.Err or the Redacted() URL.
3	Medium	runner.go:24, config.go (normalizeConfig)	An unknown adapter name or a missing required key is caught only inside the service goroutine, which logs once and returns. main keeps waiting for a signal, so restart: unless-stopped never fires. Plain yaml.Unmarshal also drops unknown keys, so a url: written at service level instead of under config: passes loading. With several services, the bad one goes quiet while the others keep running.	Call adapter.New(adapterType, cfg) for each service in normalizeConfig (factories do no I/O) and return the error, so main exits through log.Fatalf. Optionally decode with yaml.NewDecoder(...) and KnownFields(true).
4	Medium	runner.go:47, runner.go:63-66	After a successful Connect, runService never re-enters its reconnect loop, and runConnectedService logs Increment errors forever. Initialization runs only in Connect. If a provider restore or a manual drop removes the row or table, PostgreSQL and MySQL fail every tick with sql: no rows in result set or "relation does not exist", and no write lands. A dropped Couchbase collection fails the same way. Redis, Valkey and MongoDB recreate the key on their own.	Return from runConnectedService after N consecutive failures (for example 3). At line 47, return only when ctx.Err() != nil; otherwise wait reconnectDelay and loop, which re-runs New and Connect. Optionally make the increments self-seeding. PostgreSQL: INSERT ... ON CONFLICT (key) DO UPDATE SET value = keepalive.value + 1 RETURNING value. MySQL: one Exec of INSERT ... ON DUPLICATE KEY UPDATE value = LAST_INSERT_ID(value+1), then the count is 1 when RowsAffected()==1 and LastInsertId() otherwise. Do not issue a separate SELECT LAST_INSERT_ID(), which can run on another pooled connection.
5	Medium	runner.go:31	Connect gets the process-lifetime context, which has no deadline. A probe against a listener that accepts TCP and never answers showed lib/pq hanging past 6s, with or without TLS, even under a 2s context. MySQL returned at 2s. A stalled pooler handshake therefore blocks that service forever and logs nothing.	Two parts. (a) In the postgres adapter, add connect_timeout to the DSN when the user has not set one, because lib/pq v1.12.3 applies the context only to the TCP dial. (b) Wrap each Connect in context.WithTimeout, for example 60s, which is longer than Couchbase's default ready_timeout of 30s. That bounds MySQL and the initialization statements.
6	Low	adapter/redis.go:34-41, adapter/couchbase.go:101	On a failed Connect, Redis closes a.client but leaves it set, and Couchbase sets a.cluster before ensureDocument and closes it in a deferred cleanup. The runner's closeService then closes it again, logging close: redis: client is closed or close: cluster closed every 10s after the real error. This is log noise only.	Keep the client in a local variable and assign the field only on success, or set it to nil after closing.
7	Low	service_name.go:18-31	Generated suffixed names (base-N) are never recorded or checked, so they can collide with an explicit name in either order. Two services then share a log prefix.	Raise the suffix until the candidate is unused, and record every emitted name in usedNames.
8	Low	config.go:102-106, adapter/valkey.go	normalizeConfig always overwrites config.counter_key with the service-level or global value. The valkey factory ignores namespace, so a redis block copied to valkey writes the bare key counter. Neither use is documented, so the core defect is that unknown or reserved keys are accepted silently.	Reject config.counter_key in normalizeConfig. The redis/valkey merge below fixes namespace if the user approves honoring it.
9	Low	runner.go:52-70	The first Increment runs one full interval after Connect, and no adapter's Connect performs a real write once the row or key exists. With a long interval such as 24h and restarts more often than that, no write ever lands. The default 1m interval hides this.	Run one tick right after a successful Connect, before the ticker loop.
10	Low	compose.yml:15, config.go (firstExistingConfigFile), Dockerfile	A missing ./config.yml makes Docker mount an empty root-owned directory, which the loader accepts and then fails to read, so the container crash-loops. The image runs as uid 65532, so a 0600 host file is unreadable. With no .dockerignore, COPY . . puts a local config.yml and .git into the build stage (not the final image).	Reject directories in firstExistingConfigFile with a clear message. Add a .dockerignore listing config.yml, config.yaml, .env, .git and plans. Document the uid 65532 read requirement. Optionally use long bind syntax with create_host_path: false.

Adapter consolidation

Merge 1: redis and valkey on go-redis/v9 (recommended, with changes)

Both adapters run SETNX key 0 and INCR key over RESP, so one client is enough. go-redis/v9 is the better choice for four reasons.

• Size. In a scratch build (CGO_ENABLED=0, -trimpath -s -w), removing valkey-go saved 9.1 MB, while removing go-redis would save only 4.5 MB. go.mod loses one direct requirement.
• Robustness. valkey-go enables client-side caching by default and sends CLIENT TRACKING ON OPTIN on connect. Against a RESP2-only or tracking-less server it fails with ErrNoCache unless the URL carries ?client_cache=0, and adapter/valkey.go never sets DisableCache. Upstash, Garnet and older KeyDB or Dragonfly builds are the likely cases, but none was tested live. go-redis falls back from HELLO 3 to AUTH over RESP2 on its own.
• URL features. go-redis ParseURL handles redis:// and rediss:// (TLS 1.2+ with ServerName), ACL user and password, the /db path, timeouts, pool options and skip_verify.
• Existing behaviour. The namespace key and its tests already live on the go-redis adapter.

Config compatibility. Register both names on one factory: Registry["redis"] and Registry["valkey"] both map to newRedisAdapter. Generated service names come from the configured adapter string, so valkey-<host> does not change. The counter key does not change, and SETNX never overwrites, so existing counters continue.

Corrections the implementation must apply. The verifier found the first normalizer sketch too thin.

1. go-redis rejects the valkey:// and valkeys:// schemes. Rewrite them to redis:// and rediss:// with a prefix replace on the raw string, not a full url.Parse and u.String() round trip, so userinfo is never re-serialized.
2. Several query keys behave differently between the clients.
   • max_retries=0 disables retries in valkey-go but means "default 3" in go-redis, so rewrite it to -1.
   • A bare ?skip_verify (empty value) is true in valkey-go but false in go-redis. Rewrite it to skip_verify=true, or a Heroku-style URL silently turns certificate checks back on.
   • skip_verify on a non-TLS URL is ignored by valkey-go but left over in go-redis, which then fails with unexpected option. Drop it on non-TLS URLs.
   • write_timeout is applied as a dial timeout in valkey-go v1.0.75 (a copy-paste bug) and as a socket write timeout in go-redis. Keep it, because the go-redis meaning is the intended one, and note the change.
   • protocol accepts only integers in go-redis. This is minor.
3. valkey-go silently ignores unknown query keys, while go-redis rejects them. For valkey configs only, keep the shared keys (db, dial_timeout, write_timeout, protocol, client_name, max_retries, skip_verify), drop client_cache and any other unknown key, and return a clear error for addr and master_set (cluster seeds and sentinel), which have no single-client equivalent. Keep redis configs exactly as strict as they are today.
4. Keep the existing Ping in Connect. The sketch removed it, which is an unrequested change to the redis adapter.
5. Honoring namespace for valkey changes the key only for a config that already carries an ignored namespace. That is a user decision. Leave it out of the default change unless the user approves it.

Risks. A single go-redis client does not follow MOVED, so a cluster-mode endpoint that worked through valkey-go auto-detection would break. No free-tier target needs cluster mode, so an opt-in through redis.ParseClusterURL is not worth adding now. go-redis v9.23.0 is the latest release; the repo pins v9.19.0, and the bump is optional.

Hosted services this adapter covers. Redis Cloud (free 30 MB), Upstash Redis (rediss://), Aiven for Valkey, Render Key Value, Dragonfly Cloud and self-hosted Dragonfly, KeyDB, Microsoft Garnet, and self-hosted Valkey or Redis. Heroku Key-Value Store and Railway Redis are wire-compatible but have no idling free tier, so they are not keepalive targets. Fly.io's Upstash Redis is reachable only from inside Fly's private network. Koyeb does not appear to sell managed Redis (unconfirmed).

Merge 2: postgresql and mysql behind one database/sql dialect (recommended: skip)

The two adapters share a lifecycle (sql.Open, ping, CREATE TABLE IF NOT EXISTS, seed, increment in a READ COMMITTED transaction, close) and differ in driver, config key (url against dsn), DDL quoting, seed statement and read-back. A dialect struct would save about 55 to 60 net lines, not 70. It would add an indirection layer, a selectBack == "" branch and a deferred rollback with a named return, and bring no dependency or size win. The next SQL-like targets (Oracle, libSQL over HTTP) would not fit the struct cleanly. The review therefore recommends skipping the merge. If pool sizing is the actual goal, change only the SetMaxOpenConns(10) and SetMaxIdleConns(10) line in adapter/mysql.go.

If the user wants the merge anyway, keep the registry names postgresql, postgres and mysql and the config keys url and dsn. Be aware that the sketch's pool settings (SetMaxOpenConns(2), SetConnMaxLifetime(3m)) also change PostgreSQL behaviour, not only MySQL. CockroachDB silently upgrades READ COMMITTED to SERIALIZABLE unless READ COMMITTED is enabled; this is unchanged by the merge but worth a smoke test.

Hosted services these adapters cover. PostgreSQL (lib/pq): Supabase (session pooler on port 5432), Aiven for PostgreSQL, YugabyteDB Aeon (YSQL, port 5433), CockroachDB Cloud Basic, Nhost, Neon, Timescale/Tiger Cloud, Prisma Postgres (direct TCP), Render and Railway Postgres. MySQL (go-sql-driver/mysql): Aiven for MySQL, TiDB Cloud Starter, SingleStore Helios, MariaDB/SkySQL, Railway MySQL. PlanetScale is wire-compatible but dropped its free Hobby tier in April 2024.

Rejected merges

• lib/pq to pgx/v5 stdlib. The "maintenance mode" premise is stale: lib/pq shipped v1.12.0 to v1.12.3 between March and April 2026, and its README no longer carries the notice. A scratch build with pgx was 3.3 MB larger and added three indirect dependencies, for no feature this daemon uses. Note that a future lib/pq release may change the default sslmode from require to prefer, so keep ?sslmode=require in examples.
• MongoDB-compatible services. There is nothing to merge. adapter/mongodb.go is already one adapter with a mongo alias, and every compatible service connects through URI options. For the README: mongo-driver v2 requires server 4.2 or later, Cosmos DB RU and AWS DocumentDB need retrywrites=false, and DocumentDB needs tlsCAFile. The driver can move from v2.6.0 to v2.9.2 independently.
• Couchbase with anything. Its KV protocol has no sibling among the other adapters. It is the heaviest dependency (grpc, otel, protobuf, zap), but removing it means dropping Capella support, which is a user decision.
• Redis cluster or sentinel auto-detection. No free-tier target needs it, so the merged adapter returns an explicit error for addr and master_set instead.

Hosted services the unchanged adapters cover. MongoDB (mongo-driver/v2): MongoDB Atlas Free, Azure DocumentDB free tier, FerretDB Cloud, Azure Cosmos DB for MongoDB (API 4.2 or later). Couchbase (gocb/v2): Couchbase Capella free tier.

Candidate new services for the user to decide

The user decides which of these to add; nothing has been added. "Unconfirmed" means the policy could not be read on an official page, or the page leaves the key point undefined. Idle policies were checked on 2026-10-09.

Works today with existing adapters (README or example change only)

Service	Idle policy	Activity that counts	Adapter	ToS risk	Value	Source
Supabase Free	Paused after about 7 days of low activity, with an email a week before. Restorable for up to 1 year (the page anchor still says 90 days).	User database queries, API calls, Dashboard visits. Whether a direct Postgres query counts is inferred, not stated.	postgresql (session pooler port 5432, sslmode=require; direct host is IPv6-only)	Low	High	supabase.com/docs/guides/platform/free-project-pausing
MongoDB Atlas Free	Paused after 30 days of inactivity, with an email 7 days before.	Inactivity is "zero connections"; a driver write counts.	mongodb	Low	High	mongodb.com/docs/atlas/pause-terminate-cluster/
Couchbase Capella free tier	Turned off after 72 hours idle and deleted after 30 days idle.	Not defined; an SDK write is the safe choice (inferred).	couchbase (interval 12 to 24h)	Low	High	docs.couchbase.com/cloud/get-started/get-started.html
Redis Cloud Free 30 MB	Deleted permanently after 14 days with no Redis commands; no backups. Heroku listing says 30 days. Unconfirmed: the support article returns 403, so this comes from search excerpts and Redis Insight 2.64 notes.	Redis read or write commands only. Console logins do not count.	redis (daily)	Low	High	support.redislabs.com article 34131767970450
Upstash Redis Free	Archived after at least 30 days of inactivity. Data is backed up but restored into a new database, so the URL changes. One source mentions 14 days; the live FAQ says 30.	Not defined; any command should count.	redis (rediss://)	Low	High	upstash.com/docs/redis/help/faq
YugabyteDB Aeon Sandbox	Paused after 10 days of inactivity and deleted after 15 days.	Any SELECT, UPDATE, INSERT or DELETE from a connected application.	postgresql (port 5433; egress IP on the allow list)	Low	High	docs.yugabyte.com (create-clusters-free)
Aiven free PostgreSQL, MySQL, Valkey	May be powered off with "no continuative activity" or no initial use in the first few hours. No threshold is published. Powered-off services are deleted after 180 days.	Not defined; real writes are the safe reading. Whether DB traffic alone is enough is unconfirmed.	postgresql / mysql / valkey	Low to medium	High	aiven.io/docs/platform/howto/custom-plans
CockroachDB Cloud Basic	Deleted after 6 months with no activity; cannot be restored.	Not defined; SQL traffic is assumed.	postgresql (sslmode=verify-full, port 26257)	Low	Medium	docs.cockroachlabs.com (basic-cluster-management)
Azure DocumentDB free tier	Paused after 60 days of inactivity. No backups on the free tier.	Not defined (unconfirmed).	mongodb (live-check the upsert and $inc)	None found	Medium	learn.microsoft.com/azure/documentdb/free-tier
FerretDB Cloud free tier	"May be stopped or deleted if inactive"; no duration given.	Not defined.	mongodb	Low	Medium	docs.ferretdb.io/installation/ferretdb-cloud/
Nhost Starter	Project paused after 1 week of inactivity; only one active free project.	Not documented (unconfirmed whether DB traffic counts).	postgresql (public connection string)	Low to unknown	Medium	nhost.io/pricing
SingleStore Helios trial	Trial group warned after 2 days with no query, suspended 24h later, terminated 30 days after suspension. The free Shared edition has no idle rule.	Any query.	mysql (live-check the DDL dialect)	Medium (defeats a trial)	Low	docs.singlestore.com (sign-up-to-try)

Several services match an existing adapter but gain nothing from a keepalive, so they should not be listed as targets.

• Neon Free and Koyeb Postgres scale to zero after 5 minutes and keep the data. Pinging only burns compute quota (100 CU-hours a month on Neon, 5 hours a month on Koyeb).
• Render Free Postgres expires 30 days after creation whatever its activity, and Render Key Value loses data on restart.
• Railway Free sleeps services with no outbound traffic, but staying awake burns the $1 monthly credit (Free-plan forced sleep is unconfirmed).
• TiDB Cloud Starter, Nile, Prisma Postgres and Xata have no documented idle policy, or no free cloud tier.

Needs a new adapter

Service	Idle policy	Activity that counts	New adapter shape	ToS risk	Value	Source
Oracle Autonomous AI Database (Always Free)	Stopped after 7 days inactive; may be deleted after 90 cumulative days stopped. A keepalive cannot start a stopped DB.	A SQL*Net or HTTPS connection that runs SQL resets the measurement.	oracle via pure-Go github.com/sijms/go-ora/v2 (database/sql), or HTTPS POST to the ORDS REST-enabled SQL endpoint	Low	High	docs.oracle.com (autonomous-always-free)
Neo4j AuraDB Free	Paused after 72 hours without writes; deleted after 30 days paused.	Writes (from a staff forum quote of a support article).	Plain HTTP POST to Query API v2 with a single-line MERGE ... SET k.count = coalesce(k.count,0)+1	None found	High	neo4j.com/docs/aura/managing-instances/instance-actions/
Qdrant Cloud free tier	Suspended after 1 week unused; deleted after 4 weeks.	Not defined; an upsert is safest.	Plain HTTP: create a size-1 collection once, then read point 1 and upsert count+1	None found	High	qdrant.tech/documentation/cloud/create-cluster/
Zilliz Cloud Free	Suspended after 7 days of inactivity; no deletion timeline in the docs (support KB returned 403).	Not defined; use an upsert.	Plain HTTP REST v2: create collection once, then query and upsert	None found	High	docs.zilliz.com/docs/free-and-serverless-clusters
DataStax Astra DB Serverless free	Hibernated after 48 hours idle; deleted after 30 days hibernated.	Any Data API or CQL request; a request also triggers resume (503 at first).	Plain HTTP findOneAndUpdate with $inc on count and upsert: true (interval 12h or less)	None found	High	docs.datastax.com (database-statuses)
Turso Free	Archived after 10 days of inactivity; manual unarchive.	Database requests.	Plain HTTP POST to /v2/pipeline with a bearer token running an upsert RETURNING value (avoid cgo go-libsql)	Low	High	docs.turso.tech/cli/group/unarchive
CloudAMQP Little Lemur / Loyal Lemming	Deleted after two months with no publish or consume; Little Lemur also drops queues idle for 28 days.	Publishing or consuming at least one message.	amqp via pure-Go github.com/rabbitmq/amqp091-go: declare a queue, publish one message, basic.get it back	Low	High	cloudamqp.com/pricing
Aiven free Kafka	Powered off after 24 hours with no produce or consume (stated on the create page; the overview page omits it). Data is lost on power-off, and services are deleted after 180 days.	Producing or consuming messages.	kafka via pure-Go github.com/twmb/franz-go: produce one record (hourly)	Low to medium	High	aiven.io/docs/products/kafka/free-tier/create-free-tier-kafka-service
EMQX Cloud Serverless	Stopped after 30 days with no client connections; may be deleted 30 days later.	An active MQTT client connection.	mqtt via pure-Go github.com/eclipse/paho.mqtt.golang with TLS SNI: connect, publish one QoS1 message, disconnect. Needs a synthetic count or a "touch" variant of Increment.	Low	High	docs.emqx.com/en/cloud/latest/create/serverless.html
Weaviate Cloud free	Suspended after 7 days of inactivity; deleted after 30 days.	Not defined (unconfirmed).	Plain HTTP: create a vectorless collection once, then read and update one object with a fixed UUID	Low to medium	Medium	docs.weaviate.io/cloud/manage-clusters/create
InfluxDB Cloud Free	Account "may be closed" after 30 days with zero writes and zero queries.	Writes, or task and query executions.	Plain HTTP POST of one line-protocol point to /api/v2/write; read back with a Flux last()	Low	Medium	influxdata.com/influxdb-cloud-pricing-faq/
Aiven OpenSearch free	A new service with no use in 24h is powered off; the ongoing rule has no stated duration. Unconfirmed.	Indexing or querying (stated only for the first 24h).	Plain HTTP POST /<index>/_update/<id> with a painless script and upsert; shared with Bonsai	Low to medium	Medium	aiven.io/docs/products/opensearch/concepts/opensearch-free-tier
Algolia Build	"May remove Subscriber Data" after 30 days inactive; support timelines (2 months, 8 to 10 weeks) unconfirmed (403).	Searches, indexing or configuration changes (support snippets).	Plain HTTP partial update with Increment and createIfNotExists=true; the write is asynchronous	Low to medium ("evaluation purposes")	Medium	algolia.com/policies/build-plan-specific-terms
Auth0 Free tenant	All tenants scheduled for deletion after 150 days with no activity.	Dashboard logins, user logins or API calls (support article; API alone unconfirmed).	Plain HTTP: client-credentials token, then PATCH /api/v2/clients/{id} metadata	Low to medium	Medium	auth0.com/docs/secure/data-privacy-and-compliance/data-processing
GitHub Actions schedules (public repos)	Scheduled workflows disabled after 60 days with no repository activity.	Not defined; commits are commonly understood to count.	Plain HTTP PUT .../actions/workflows/{id}/enable with a fine-grained PAT	Medium	Medium	docs.github.com (disable-and-enable-workflows)
Render free web services	Spin down after 15 minutes without inbound traffic; 750 instance hours a month.	Inbound HTTP requests and WebSocket messages.	Generic HTTP GET with a 60 to 90s timeout	Medium (AUP "bypass usage restrictions"; quota drain)	Medium	render.com/docs/free
Bonsai Sandbox	Deleted after "several weeks" unused (2023 docs). Free plan now appears only on Heroku. Unconfirmed.	Any incoming traffic.	Shared OpenSearch/Elasticsearch HTTP adapter	Low	Low	bonsai.io/docs (periodic-cluster-cleanup)
Upstash Vector	No Vector-specific policy; the 30-day archive rule is documented only for Redis. Unconfirmed.	Not defined.	Plain HTTP fetch and upsert	None found	Low	upstash.com/docs/redis/help/faq
Hugging Face Spaces (cpu-basic)	Sleeps after 48 hours inactive; new free CPU Gradio and Docker Spaces now need a paid plan.	Visitors (whether a bare GET counts is undocumented).	Generic HTTP GET (24h or less)	Low to medium	Low	huggingface.co/docs/hub/spaces-gpus
Koyeb Free Instance	Scales to zero after 1 hour; new signups need a paid plan since February 2026.	Internet traffic over HTTP/1.1 (HTTP/2 cannot wake it).	Generic HTTP GET	Low to medium	Low	koyeb.com/docs/run-and-scale/scale-to-zero

The following were checked and are not worth an adapter.

• No idle policy: Cloudflare D1, KV and R2; Azure Cosmos DB for NoSQL; Pinecone Starter (third-party pause claims are unconfirmed); Firebase Spark; Convex; Synadia; Docker Hub; MotherDuck; Chroma; Tiger Cloud.
• Trial or time-based expiry: Elastic Cloud, ClickHouse Cloud, Confluent Cloud, Meilisearch Cloud, Typesense Cloud, ScyllaDB Cloud.
• Discontinued or ending: Glitch, Upstash Kafka, HiveMQ Serverless (ends 2026-12-31).
• Policy keyed to something a remote write cannot reach: Appwrite (Console activity only), PythonAnywhere (owner confirmation), Streamlit Community Cloud (WebSocket visits and a manual wake button), Google Colab (UI only, automation prohibited), Oracle Always Free compute (on-host CPU, memory and network use).
• Cost features, not idle policies: Fly.io autostop, Railway app sleeping, Replit Autoscale, PocketHost hibernation (a few seconds), Google Cloud Pub/Sub subscription expiry (set it to never expire instead).

Proposed new adapters ranked

The ranking weighs how many confirmed, high-value services an adapter unlocks against the dependency it adds. All stay pure Go.

1. http (generic, plain net/http, no new dependency). One configurable request (method, URL, headers, body, expected status) covers Render web services, Hugging Face Spaces, Koyeb, GitHub workflow enable, Auth0 (with a token step), Oracle ORDS and Nhost GraphQL. It has the widest reach and the lowest cost. The current interface expects Increment to return a counter, so this adapter needs a local tick count or a "touch" variant of the interface. A JSON path for reading a counter back would let it serve some database APIs too, but that adds config surface.
2. Thin HTTP-database adapters sharing one helper. Qdrant, Zilliz, Neo4j Query API, Astra Data API, Turso pipeline, Weaviate and InfluxDB each need a setup call, an increment-like write and a JSON response parse. They share auth headers, a JSON client and error handling. Either write one small adapter per service on a shared httpjson helper, or treat them as presets of the generic http adapter. Qdrant, Neo4j, Astra and Turso have the best-confirmed policies.
3. oracle via github.com/sijms/go-ora/v2. Oracle's Always Free database has an explicitly documented reset condition and a hard 90-day deletion risk. It follows the existing SQL adapter pattern. Use a PL/SQL block that tolerates ORA-00955 for table creation, then MERGE and UPDATE ... RETURNING. Wallet-less TLS needs mTLS set to "not required" with an ACL.
4. amqp via github.com/rabbitmq/amqp091-go. It covers CloudAMQP (RabbitMQ and LavinMQ) and any AMQP 0-9-1 broker, with a confirmed policy and a small dependency.
5. kafka via github.com/twmb/franz-go. It covers Aiven free Kafka, whose 24h power-off loses data, and any Kafka-compatible broker. The dependency is moderate in size.
6. mqtt via github.com/eclipse/paho.mqtt.golang. It covers EMQX Serverless and self-hosted brokers. Its value is lower because the idle window is 30 days and HiveMQ's free plan is ending.

Unresolved questions

1. Should the merged valkey adapter honor namespace? Doing so changes the key for any valkey config that already carries an ignored namespace.
2. Is any deployment pointing adapter: valkey at a cluster-mode or sentinel endpoint? The merge would break it, and the answer decides whether cluster support is needed now.
3. Should the Adapter interface gain a "touch" operation without a counter, which the generic http, mqtt and kafka adapters need, or should those adapters return a local count?
4. Should normalizeConfig reject unknown YAML keys (KnownFields(true))? That would be stricter for existing configs that carry stray keys.
5. Does plain database traffic prevent Aiven's free power-off, or is a Console login also needed? Only a live test can answer this.
6. Should README examples for Couchbase, Astra and Neo4j recommend a shorter interval, given their 48 to 72 hour windows, and should Supabase docs warn against the 6543 transaction pooler with lib/pq?
7. Is the user willing to accept the Medium ToS risk of keeping Render web services or GitHub scheduled workflows alive, or should the generic http adapter be documented for warm-start use only?