ash_onetime
ash_onetime is an Ash extension for explicit keyed-effect semantics. It separates
replay-safe idempotency from collision-rejecting one-time nonces and uses PostgreSQL as the
authoritative admissions store.
What it does
Every protected action declares one strategy and a nonempty scope. PostgreSQL admits exactly once per key under the declared scope; the unique constraint — not application code — decides concurrent races. There is no admission pre-read.
:idempotency— execute the effect once per key, store the classified response, and replay it verbatim on retries within the retention boundary. A conflicting fingerprint (same key, different request) is terminal and never re-executes. The effect, the admission claim, and the encoded response commit or roll back together in the action's transaction.:one_time_nonce— authenticate a single opportunity and reject every collision. The spend lives and dies with the action transaction by default; a retry must bear a fresh proof.commit: :independent(nonce, opt-in, RFC 9449 §11.1) — the DPoP replay fence. The nonce claim commits in its own transaction before the action body runs, so a downstream failure cannot make the proof reusable. A reused proof is rejected with:nonce_already_usedfor the acceptance window. Default-off; existing nonce consumers are byte-for-byte unchanged.
Optional, non-admitting surfaces: a response cache, a Plug, Oban workers for cleanup and forward partition creation, and an external-effect execute/recover protocol for peers that need to observe or reverse a side effect.
Hosts that already own one authoritative Ecto transaction can use the public
AshOnetime.Transaction boundary instead of wrapping an Ash action. It reserves idempotency
and nonce authority inside the caller's existing READ COMMITTED transaction, so the claim,
the host effect, and the exact replay bytes commit or roll back together. Logical partitions
isolate tenants or authority planes without importing the host's actor types. See
Transaction-owned admission.
When to use this vs. hand-rolled idempotency
Use ash_onetime when you need a correct admission layer for effectful Ash actions and do
not want to re-derive the failure modes a hand-rolled table-plus-flag approach ships with.
- The correctness guarantee: a PostgreSQL-authoritative, once-per-key local effect and typed replay within the declared retention boundary. The effect, the admission claim, and the encoded response commit or roll back together in the action's existing transaction; there is no admission pre-read, so the unique constraint — not application code — decides concurrent races. A conflicting fingerprint is terminal and never re-executes.
- What it replaces: the bespoke idempotency-key table, the "did this already run?" pre-check that races under concurrency, the stored-response schema you have to version and re-encode, the fingerprint binding you have to remember to update, and the silent double-execution that lands when you forget. Each of these is a class of bug this library was built to make impossible at the boundary.
- When NOT to use it: read-only or non-transactional actions (the library rejects them), resources without AshPostgres (PostgreSQL is the authoritative store; there is no fallback), and workloads that need end-to-end exactly-once delivery rather than once-per-key local admission (delivery is a separate concern — see External effects). If you have no effectful action to protect, there is nothing to admit.
Quick start
defmodule MyApp.Charge do
use Ash.Resource,
domain: MyApp.Domain,
data_layer: AshPostgres.DataLayer,
extensions: [AshOnetime.Resource]
postgres do
table "charges"
repo MyApp.Repo
end
attributes do
uuid_primary_key :id
attribute :account_id, :uuid, allow_nil?: false, public?: true
attribute :amount, :integer, allow_nil?: false, public?: true
end
actions do
defaults [:read]
create :charge do
transaction? true
argument :idempotency_key, :string, allow_nil?: false
accept [:account_id, :amount]
end
# A DPoP-protected redemption: the proof's spend survives a downstream failure.
action :redeem, :integer do
transaction? true
argument :value, :integer, allow_nil?: false
argument :proof, :string, allow_nil?: false
run MyApp.Redeem
end
end
onetime do
protect :charge do
strategy :idempotency
scope([{:static, "charge"}, {:attribute, :account_id}])
key({:client, :idempotency_key})
fingerprint(attributes: [:account_id, :amount])
response(MyApp.ChargeCodec, fields: [:id, :account_id, :amount], classify: MyApp.Classifier)
retention({24, :hour})
end
protect :redeem do
strategy :one_time_nonce
scope([{:static, "redeem"}])
key({:verified, :proof, MyApp.ProofVerifier})
window(max_age: {5, :minute}, clock_skew: {30, :second})
commit :independent # RFC 9449 §11.1 replay fence (omit for the default :with_action)
end
end
end
Install with mix ash_onetime.install (Igniter-powered) — it wires the repo, generates the
migration, and sets up the schema. See the
Getting started guide for the full walkthrough.
Try it
Runnable Livebook notebooks — one per concern — walk through each strategy against a real
PostgreSQL. Open them in Livebook, set DATABASE_URL, and run every
cell. Each notebook's code is regression-pinned so it never ships broken.
- Idempotency — fresh execution, replay, fingerprint conflict.
- One-time nonces — spend, reuse rejection, and the
commit: :independentreplay fence. - External effects and recovery — the execute/recover protocol.
See Security model for the authority and fail-closed contract, and Recipes for end-to-end payment, webhook, and redemption patterns.
Status
The current package release is v1.2.0 on Hex. It adds
the mix ash_onetime.doctor --live schema-currency preflight, the backup/restore runbook,
and constant-time digest-comparison unification on top of v1.1's transaction-owned admission
and logical partitions. The
source is public. Every protected action chooses
:idempotency or :one_time_nonce and declares a nonempty scope; there is no default
strategy or global scope fallback. PostgreSQL-authoritative admission, transactional Ash
execution, typed replay, fail-closed nonce spending, signed tokens, external-effect recovery,
bounded cleanup, optional cache/Plug/Oban integrations, the DPoP replay fence, and release
gates are present.
Compatibility
- Elixir
~> 1.20(developed and tested on 1.20.2) - Erlang/OTP 29
- Ash
>= 3.31.3and< 4.0.0(the whole 3.x line from the 3.31.3 floor up) - AshPostgres 2
- PostgreSQL 18 for the project test harness. The SQL surface requires PostgreSQL 11+
(declarative hash/range partitioning with default partitions,
SKIP LOCKED,pg_advisory_xact_lock(bigint)); versions below 18 are not exercised by this project's CI — treat them as unverified.
The floor is Ash 3.31.3, not the earlier 3.x line: EEF-CVE-2026-55736 (private
action arguments settable by user input, fixed in 3.29.3),
EEF-CVE-2026-70395 (predicate injection in manage_relationship belongs_to
lookup disclosing secret lookup keys, fixed in 3.31.1), EEF-CVE-2026-69659
(memory exhaustion via unbounded keyset-cursor deserialization, fixed in
3.31.1), and EEF-CVE-2026-67579 (filter expression injection via a forged
keyset pagination cursor — HIGH, fixed only in 3.31.3) all affect Ash below
3.31.3 — a security library must not admit a vulnerable floor. Compatibility
across the range is verified per matrix cell by the standard gate battery — format,
compile with warnings-as-errors, mix hex.audit (Hex security advisories),
mix deps.audit, the full test suite, mix credo --strict, mix dialyzer,
mix docs --warnings-as-errors, and mix hex.build — run against the 3.31.3 floor
and the latest published Ash 3.x. The release battery (mutation matrix,
unpacked-package check, DSL cheat-sheet freshness) runs once per push in the
release-checks job against the committed lock, not per cell.
.github/workflows/ci.yml is configured to re-run this matrix on every push and
pull request. The pinned development runtime is Elixir 1.20.2 / Erlang/OTP 29
(.tool-versions).
Development
Start the dedicated test database and run the suite as documented in CONTRIBUTING.md, then:
mix deps.get
DATABASE_URL=ecto://postgres:postgres@127.0.0.1:18841/ash_onetime_test mix test
See CONTRIBUTING.md for the complete gate battery and usage-rules.md for non-negotiable integration boundaries. The Getting started guide covers installing the package and protecting your first action.
Handling results at the call boundary
A protected action's failure carries a typed :code that survives the Ash pipeline, and its
success carries a replayed-vs-fresh signal. Both are observable after Ash.create/2 /
Ash.run_action/2 returns:
case Ash.create(changeset) do
{:ok, record} ->
# 201 on fresh execution (replayed? == false), 200 + Idempotent-Replayed on retry (true).
status = if AshOnetime.replayed?(record), do: 200, else: 201
{:error, error} ->
# The typed code reaches the caller — map it to HTTP.
case AshOnetime.Error.code(error) do
:nonce_already_used -> {:conflict, "nonce was already used"}
:key_reused_with_different_request -> {:conflict, "key reused with a different request"}
:request_in_progress -> {:conflict, "request is already processing"}
nil -> {:internal_server_error, "unexpected error"}
end
end
AshOnetime.replayed?/1 is tri-state: true (tracked replay), false (tracked fresh), or
nil (untracked execution, primitive-return action, or unprotected — see
Replay). The full code→HTTP table is in
Errors.
Guides
- Resource DSL
- Idempotency
- One-time nonces
- External effects and recovery
- Transaction-owned admission
- Replay: fresh vs stored
- Errors and HTTP mapping
- Operations
- Security model
- Recipes
- Phoenix integration
- Telemetry
- Upgrading
- FAQ
- Livebook notebooks: Idempotency · Nonces · External recovery
- Generated Spark DSL reference
License
MIT. See the repository LICENSE file.