Edge Config Store: TTLs and Point-in-Time Reads

Edge Config Store: TTLs and Point-in-Time Reads

You've joined the platform team that owns the edge config store: a small versioned key-value store that CDN edge nodes read feature flags, service URLs and sessions from. Every write is kept as a version, which today only powers the history debugging endpoint. There is one bug ticket and two feature tickets in your queue.

Codebase

How your code is run

run(operations: list[list]) -> list runs the operations in order against a fresh store and returns one result per operation. Every operation is [name, ts, ...args]; ts is an integer number of seconds and never decreases (several operations may share a ts). Errors are returned as "error: <code>" strings, never raised.

• Keys are case-sensitive, non-empty strings of at most 200 characters with no whitespace. Values are strings (empty allowed). Arguments are validated in the order listed.
• A key is live if its latest write is a set (not a delete).
• Multiple writes at the same ts are applied in operation order; the last one wins.
• Unknown operation → "error: unknown operation". Wrong number of arguments or non-integer ts → "error: invalid arguments".

Worked example

[["set", 1, "user:42:name", "Ada"],
 ["set", 2, "user:7:name", "Grace"],
 ["set", 3, "flags:dark_mode", "on"],
 ["get", 4, "user:7:name"],
 ["delete", 5, "flags:dark_mode"],
 ["delete", 6, "flags:dark_mode"],
 ["size", 7],
 ["history", 8, "flags:dark_mode"]]

returns [true, true, true, "Grace", true, false, 2, [[3, "on"], [5, null]]].

Part 1 — scan returns the wrong rows (EDGE-311)

Goal: make scan match its contract.

Edge nodes call scan to load a namespace of config at boot. Two reports from the edge team: "the rows come back in a strange order, so our diff-based sync thinks everything changed", and "scan("user:") returned our legacy User: and USER: keys too".

Rules

1. scan returns only live keys that start with prefix, compared case-sensitively. The empty prefix matches every key.
2. Rows are ordered by key, ascending, using plain string comparison (so "Zulu" < "alpha").

tests/test_store.py::test_scan_is_ordered_by_key currently fails.

Part 2 — Key expiry (TTL) (KV-88)

Goal: sessions and temporary flags should expire on their own.

Rules

1. ttl must be an integer >= 1 (not a boolean), otherwise invalid ttl and nothing is written.
2. A key set at ts with ttl is live for ts <= t < ts + ttl. At t = ts + ttl it is gone (exclusive boundary). The remaining seconds reported by ttl at time t is ts + ttl - t.
3. set without a ttl writes a value that never expires (it replaces any earlier expiry).
4. expire gives a live key the new expiry now + ttl, replacing any earlier one (it may shorten it).
5. An expired key behaves exactly like a deleted one for get, mget, scan, size, delete (returns false), expire (returns false) and ttl (returns null), until it is set again.
6. history is unchanged: rows are still [written_ts, value]; expire and expiry itself add no rows.

Example

[["set", 100, "session:abc", "u-42", 30],
 ["get", 129, "session:abc"], ["ttl", 129, "session:abc"],
 ["get", 130, "session:abc"], ["ttl", 130, "session:abc"]]

returns [true, "u-42", 1, null, null].

Part 3 — Point-in-time reads (EDGE-340)

Goal: let incident responders ask "what did the edge see at time T?".

Rules

1. at must be an integer (not a boolean) with at <= ts, otherwise invalid timestamp. at may be earlier than any write (the answer is then null / []).
2. The answer reflects every write made at or before at (including writes made at exactly at), deletes, and expiry as of at: a version that had expired by at yields null; it never falls back to an older value.
3. An expire issued at time E never changes what a read at a time before E returns.
4. Historical reads must not change the result of any later get, scan, size or ttl.
5. All Part 1 and Part 2 rules apply to scan_at / get_at (ordering, case-sensitive prefixes, exclusive expiry).

Example

[["set", 10, "feature:checkout", "v1"],
 ["set", 20, "feature:checkout", "v2"],
 ["delete", 30, "feature:checkout"],
 ["get_at", 40, "feature:checkout", 15], ["get_at", 40, "feature:checkout", 20],
 ["get_at", 40, "feature:checkout", 35], ["get_at", 40, "feature:checkout", 5]]

returns [true, true, true, "v1", "v2", null, null].

Notes

• history.py and batch.py are known-good. The Version.expires_at field was added ahead of Part 2 but has never been exercised.
• You may use the AI assistant freely; verify what it gives you.
• Later parts must not break earlier ones; hidden tests re-check earlier behaviour, including randomised workloads that mix every command.