Roadmap
v0.1.0 — first release · implemented
v0.1.0 was built and tested in three stages. Each stage had to pass its tests in Studio before the next one began.
| Stage | Modules | Verified by |
|---|---|---|
| 1. Foundation | Core/App, Signal, Cleanup, State, Random | Unit tests, demo place, examples |
| 2. Secure infrastructure | Network, Component, Pool, Zone, Debug | Unit tests, plus a Play-mode test with a hostile client against real remotes |
| 3. Advanced, optional | Promise, Data (beta in 0.1.0 and 0.2.0), Replication, State UI bindings | Data failure simulations (throttling, takeover by another server, migration failures, newer data, leaving mid-load, save failures, shutdown budget), plus a Play-mode Replication test |
Scope notes:
- Data uses adapters, with ProfileStore first. A custom DataStore backend would only be added after its session-locking, throttling, shutdown-budget and migration-failure simulations pass. Data must never overwrite stored data after a failed load.
- Replication only sends what the server explicitly shares, and only from server to client.
v0.2.0 — verification and optimization · released 2026-11-14
This stage proved that what exists works well. No new modules were added.
- Data: ✅ tested against real ProfileStore 1.0.3 in its mock mode, including save completion and shutdown.
- Pool: ✅ instance reuse benchmarked (1.7–2× faster than clone and destroy).
- Benchmarks: ✅ see Benchmarks. They led to a 3–4× faster Signal, a further 3–5% on
Firefrom inlined dispatch, and 4× less Cleanup memory. - Strict type check: ✅ clean under luau-lsp with Roblox types, and enforced in CI along with StyLua and selene.
- Random: ✅ chi-squared tests for the weights, the Luck formula and shuffle; a server-authoritative reward pattern is documented.
- Reviews: ✅ three rounds of review fixes, including a Data shutdown race and runaway
Startthreads; every fix has a regression test (see the changelog).
v0.3.0 — types, Data out of beta, hardening · released 2026-11-14
Still verification work. No new modules.
- Data: ✅ out of beta. Every CI run tests it on real DataStores on real Roblox servers: a player moving to a second server (two servers at once), many sessions at once, and the live DataStore tests. Studio Server & Clients with 5 players covered players leaving together and during shutdown. Still to observe: a live game under heavy load over time.
- Types: ✅ typed public API (
Promise<T>,Store<T>/Profile<T>, realselftypes). - Hardening: ✅ cancelled waiting threads are skipped everywhere, and settled promises, disconnected connections and players who left mid-load no longer stay referenced. Each fix has a regression test.
- Performance: ✅ fewer allocations in Random, Promise, State, Cleanup, Network, Replication and Zone, measured A/B against 0.2.0 on a Roblox server (see Benchmarks).
- Packaging: ✅ Wally and pesde packages checked in CI; 0.3.0 is published as
layfate/constellaon Wally andzense/constellaon pesde. Fresh installs from both registries match every library module.
v0.3.1 — lifecycle, concurrency and Delta replication · released 2026-11-14
- Lifecycle: ✅ cancelled Data loads, Component
Startthreads, Signal runners and Pool prewarm failures clean up fully. - Network: ✅ per-player and overall in-flight request limits (16 and 256 by default).
- Promise: ✅ executors and handlers on recycled threads: about 2× faster and up to 56% less memory per promise.
- Replication: ✅ one validation walk (full sends a third faster) and the opt-in Delta mode (95% fewer bytes for small changes to big tables by estimate; 89% measured on the wire with real clients in 0.3.7).
- Zone, Network, Signal: ✅ fewer allocations and checks per step, request and connection.
- Benchmarks: ✅ every module, a 50-player tick, LemonSignal/Janitor/evaera Promise as competitors, confidence intervals, and a CI gate that fails a pull request with a regression.
v0.3.2 — departed players and faster hot paths · released 2026-11-14
- Network, Replication: ✅ requests and shares for players who already left keep nothing alive; refused Replication resyncs are retried with backoff.
- Data: ✅
profile:UpdatefiresChangedfor in-place edits. - Performance: ✅ nested Signal
Fireand promise chains reuse parked runner threads (25–34% faster, no garbage per nesting level); StateSet32% faster; Network requests allocate nothing.
v0.3.3 — faster payload checks · released 2026-11-16
- Replication: ✅ payload checks add plain keys and values inline: full sends 32% faster and Delta sends 28% faster on a Roblox server, with identical sizes and rejection reasons.
- Packaging: ✅ wally.run links to the documentation; the README covers exported types with Wally.
v0.3.5 — starter template and typed demo · released 2026-11-17
- Template: ✅ a starter game with the recommended layout across ReplicatedStorage, ServerScriptService, ServerStorage and StarterPlayerScripts, checked in CI and in a Studio playtest with a hostile client. See Game layout.
- Demo: ✅ typed with Constella’s exported types instead of
any.
v0.3.6 — every public type from the root module · released 2026-11-17
- Types: ✅ all 71 public types are exported from the root module (
Constella.Context,Constella.Store<T>, …), so Wally and pesde users can use them. Lazy loading is unchanged and tested.
v0.3.7 — real clients and measured traffic · released 2026-11-17
- Real clients: ✅ load tests with 1–8 real Studio clients over the network replicator; every replica matched the server.
- Bandwidth: ✅ measured from the engine’s counters: Delta sends 130 bytes against 1,184 for the full value (89% less).
- Signal: ✅ 144 bytes per connection instead of 160, compared like-for-like with sleitnick/signal (which is still faster).
- Soak: harness shipped; a multi-hour result is not published yet.
Next
- Soak: publish a multi-hour soak result with its control run.
- Pool: measure memory over long sessions.
- Replication Delta: run the Play-mode test (
tests/client/ReplicationIntegration.client.luau) in Studio with real clients. - Typed listeners and profile keys:
Cleanup:Connecttypes its callback as(...any), andprofile:Setaccepts any key and value.
Toward 1.0
- API freeze, then semantic versioning for every breaking change.