Roblox Performance and Release

Target: Roblox Studio (October 2026 build) on macOS, driven through the scenario-roblox-expert toolkit.

Expert level here means every claim is a number from the side that hurts, taken before and after one change, and every fix removes work instead of hiding it. Budgets come from the low-end phone, and Studio numbers are labeled as Studio numbers. Release is part of performance: ship with a restart plan and a rollback path, and never publish without the owner.

If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-roblox-expert (channels, review loop, toolkit, current-Roblox traps).

Stance (the expert delta)

1. Budget for the low-end phone; label Studio numbers as Studio numbers. McNeill's budget card targets a 2 GB phone: 500 draw calls, 500,000 triangles, 1.3 GB, 150,000 instances, 50 KB/s down (RDC 2024). Phones run CPU work about 4 times slower than a desktop. Studio memory "is incorrect in Studio, you need to test live" (McNeill). Studio runs the playtest Server and Client in one process, so server work shows in client frames [added, measured].
2. Name the cause before the fix. Label every per-frame system with debug.profilebegin and fix the longest lowest-level label (Roblox Learn, MicroProfiler). Studio's execute_luau reaches the Script Profiler, so an agent gets the same answer by code: BasePart.CFrame [C] and the scan label led the lab's server time (measured).
3. Gate on frame consistency, compared per frame. Fifty-nine 10 ms frames plus one 410 ms frame stutter at "60 fps" (Roblox Learn). Use the lead gate: p95 at most 1.2 times the budget, worst at most 2 times; healthy places read p95 17.2 to 17.4 ms at 60 fps. Compare script time per frame: per-frame work runs more often once a fix raises the frame rate [added, measured].
4. Remove server work players never see. "Never tween on the server" (Crusherfire). Resizing a welded server part re-replicates the whole assembly (McNeill and Lynn, RDC 2025). Anchoring idle zombies halved one game's server CPU (same talk). Anchoring is not enough when the server still moves the parts: anchored coins spun by the server sent 7.6 to 10 MB/s in Studio, and 12 KB/s once clients spun them (measured; DataSendKbps is in kilobytes per second).
5. A leak is a rising floor, named by its owner. Healthy memory is a sawtooth (Roblox Learn). Snapshot 3 or 4 times in the same state and trust only rows that keep rising (Crusherfire). Server leaks matter most, because a server outlives hundreds of players (Crusherfire). Destroy does not release a part's connections to other signals: each orb stayed in memory, unparented, held by its Heartbeat closure (measured).
6. Replication is memory, bandwidth and join time. Tags and attributes on every part replicate to every client: one game's crash rate rose from 3 to 4 percent to 45 percent (McNeill and Lynn). Each server-owned moving assembly costs about 1 KB/s, and a join replicates about 10,000 instances per second (McNeill). Keep server-only data in Luau tables and heavy folders in ServerStorage; tune streaming instead of turning it off (Roblox Learn).
7. Release is part of performance, and a person ships. Handle ServerRestartScheduled, restart only outdated servers, and investigate a client crash rate above 2 to 3 percent after an update (docs). A rollback is restore, publish, restart: restoring alone publishes nothing (docs). Agents prepare and dry-run; the owner publishes.

Establish first

• Symptom and side: low client frame rate, server lag (heartbeat under 60, rubber-banding for everyone), crashes or memory, join time, or bandwidth. One owner per symptom.
• Target devices: the lowest phone you support. Default: a 2 GB phone, 60 fps, budgets in critique.md.
• Live data: Performance dashboard numbers (crash rate, server memory by age, frame rate by device) if the game is live.
• Repro: a place or a scripted scenario that shows the problem in a playtest, with a debug spawner for load (N players, N NPCs, N items).
• Release state: unreleased, live with players, or mid-update; who may publish and restart (always a person).
• Defaults when unanswered: measure the Server and the Client, change one thing per playtest, no publish.

Procedure

1. Reproduce under load. Goal: the slow case in a playtest, with numbers on both sides. How: a debug spawner driven by Workspace attributes (N players, N NPCs, N items; McNeill and Lynn), then rbx_perf.measure(pt, 6) on Server and Client at the same server time() (P2). GATE: baseline saved for both sides; the side that fails the frame gate is named.
2. Audit statically. Goal: costly patterns and counts before playtesting. How: rbx_perf.audit_place(path) in Lune; PerfAudit.all(game) in Edit (P1). GATE: every warn fixed or justified in one line; counts inside the Numbers table.
3. Name the cause. Goal: the function, label or asset that owns the cost. How: rbx_perf.script_profile(s, side) in the playtest; MicroProf.capture for labels in Edit; SceneScan.all() for triangles, instances and script memory (P3, P6). GATE: the top function or label is named with its ms per frame.
4. Fix one thing per playtest. Goal: a proven gain, nothing else changed. How: flip one attribute, measure at the same offset, rbx_perf.compare(before, after) with its 5 percent noise floor. Patterns in P4. GATE: the metric the fix touches improves past the noise floor; no other metric gets worse; the frame gate holds on the side that hurts.
5. Soak for leaks. Goal: memory that never comes down, with its owner. How: PerfProbe.soak, HeapProf.diff and SceneScan UnparentedInstances on the Server over 20 s or more, leak path on and off (P5). GATE: memory category growth and unparented instances stay flat; tracked objects stay bounded.
6. Check scene, draws and streaming. Goal: the worst view fits the phone budget. How: Client Stats.SceneDrawcallCount and SceneTriangleCount, the batching probe (P7), streaming radii in the project file (P8). GATE: draw calls, triangles and instances inside the Numbers table at the worst view.
7. Hand the phone list to the owner. Goal: device numbers Studio cannot give. How: the list in P9 and the GUI paths in gui-paths.md. GATE: memory and frame rate from a real phone, or the report says "Studio only".
8. Release. Goal: an update players survive, with a way back. How: the restart handler (P10), saves from scenario-roblox-data, ShipCheck from scenario-roblox-tooling; the owner publishes and restarts outdated servers. GATE: handler tested in a playtest; the owner's go recorded; the agent never publishes.

Numbers

Value	Number	Relative to
Scene draw calls	500 or fewer, UI included	low-end phone, worst view (McNeill)
Scene triangles	500,000 or fewer	same
Shadow draw calls	up to 3 times the scene draws	same
Client memory	under 1.3 GB; ejections start past about 66 percent of 2 GB	2 GB phone, on device (McNeill)
Instances	under 150,000; about 10,000 per second on join	join time (McNeill)
Replication down	under 50 KB/s	per client (McNeill; Crusherfire)
Server-owned moving assembly	about 1 KB/s each	McNeill
Phone CPU	about 4 times a desktop	10 ms desktop is about 40 ms
Frame gate	p95 at most 1.2 x budget (20 ms), worst at most 2 x (33.3 ms); healthy places read p95 17.2 to 17.4 ms	60 fps (lead gate)
Server memory	6.25 GiB plus 100 MiB per player; keep under half	docs
Client crash rate	investigate above 2 to 3 percent	docs
BulkMoveTo	0.255 ms against 0.49 ms of per-part writes	Crusherfire's spinner
Lab, all fixes	Server script 12.0 to 0.36 ms per frame; server sends 6,526 to 11 KB/s; HUD instances 6,002 to 302	Studio 0.743, desktop, one process

Quality gates

Measurable:

• Frame gate on the side that hurts, from PerfProbe.window(...).gate or the lead Stats.frames.
• rbx_perf.compare rows past the noise floor for the metric each fix touches, with one result file per variant.
• PerfAudit 0 warn findings, or each one justified.
• Soak: memory category growth and unparented instances flat; tracked objects bounded.
• SceneScan and Client Stats: draw calls, triangles and instances inside the Numbers table.
• Every per-frame system carries a debug.profilebegin label; long-lived threads set debug.setmemorycategory.
• Restart handler connected in a playtest; P9 phone numbers reported by the owner or marked "Studio only".

Visual: a playtest capture through rbx_review of the worst view before and after (rbx_review.contact_sheet), and the MicroProfiler or Script Profiler capture taken by hand when a claim cites it (gui-paths.md). The full rubric is in critique.md.

Common mistakes

Mistake	Symptom	Fix
Gating on fps or Studio memory	Numbers that phones never show	Frame-time p95 and worst; the P9 phone list
Comparing ms per second after a fix	Per-frame work looks worse once fps rises	Compare ms per frame
Gating on Stats.HeartbeatTime, Client DataReceiveKbps or FrameTime in Studio	Swings of 8 to 100 ms; 0 KB/s; 323 ms frames	Heartbeat dt, Server DataSendKbps, physics counts, Script Profiler
Two changes in one playtest	No change can be credited	One attribute per run
Windows taken at different times	Leaks make later runs heavier	Wait for the same server time()
Anchoring but still moving parts on the server	MB/s of replication	Move cosmetics on clients
Destroy without releasing outside connections	Unparented parts and table entries climb	Disconnect in Destroying:Once
Reading gcinfo() from execute_luau	About 0.9 MB: the injected chunk's heap	HeapProf memory categories
Server tweens or resizing welded parts	Replication spikes	Client-side tweens and resizes
Tags or attributes on every part	Memory and crash rate rise	Server data in Luau tables
Reading render stats from a covered Studio window	Counters frozen at one value	Put parts in view and check the counters move first
Grid cells keyed by Vector2	Every lookup misses	Vector3 or number keys
Publishing to test a fix	Players get an unproven build	Owner publishes; Version History restore, publish, restart

Handoffs

• Receives from scenario-roblox-environment: EnvAudit counts and per-model streaming modes (environment owns per-model modes; streaming settings are decided here).
• Receives from scenario-roblox-luau: Bench and Parallel results, native candidates and leak slopes.
• Receives from scenario-roblox-gameplay: NPC bench numbers; from scenario-roblox-vfx-audio: emitter, beam and flipbook counts; from scenario-roblox-networking: payload sizes and send rates.
• Delivers to scenario-roblox-luau: the leaking function or hot loop. Cleanup code (Leak, class template), native code and Actors live there.
• Delivers to scenario-roblox-tooling: the go for a release. ShipCheck, Saved versions, Open Cloud publishing and :restartServers automation live there; agents run dry runs only.
• Delivers to scenario-roblox-data: the restart plan, so BindToClose and session saves fit the 30 s shutdown budget.
• Delivers to scenario-roblox-ui, scenario-roblox-assets and scenario-roblox-animation: UI draw calls, texture or mesh budgets and rig timers (updateInvalidatedFastClusters) that failed a gate.

Current Roblox notes

• ScriptProfilerService and HeapProfilerService are PluginSecurity. Studio's execute_luau reaches both on the playtest Server and Client; a game Script gets "lacking capability Plugin" (measured, P3).
• SceneAnalysisService exposes the Scene Analysis window as six queries. It answered in Edit, on the playtest Server and on the Client, and a game Script can call it (measured).
• Draw calls and triangles are script-readable on the Client through Stats.SceneDrawcallCount and SceneTriangleCount. They froze when the Studio window stopped drawing 3D (measured).
• Streaming is on for new places. Its Workspace settings are NotScriptable for game scripts and absent from the type definitions. From execute_luau only StreamingEnabled reads and writes; the radii fail with "is not a valid member of Workspace" (measured). Set them in the Properties panel or the Rojo project.
• RunService.PreRender replaces RenderStepped. Lighting.LightingStyle replaces Technology (Future is Realistic plus PrioritizeLightingQuality). 4K textures are live since 2026-01-30.
• Publishing needs the Maturity and Compliance questionnaire, and the place publishing API cannot update SurfaceAppearance, unions, EditableImage, EditableMesh or BaseWrap content.
• require("@rbx/LibMP") exposes the MicroProfiler to execute_luau and the command bar (MCP skill rbx-perf-profiling). Edit snapshots hold full frames; playtest snapshots held 0 to 128 frames, fewer on heavy scenes (measured).

References

• references/expert-notes.md: principles and judgment by expert, with timestamps and live measurements.
• references/procedures.md: full code for stages P1 to P10, each with its test id and live result.
• references/critique.md: pass and fail rubric and the budget gate table.
• references/memory-and-crowds.md: client memory composition, asset and batching audit (PerfAudit.assets), character and pet cost (updateInvalidatedFastClusters over 4 ms), crowds that streaming cannot help.
• references/gui-paths.md: MicroProfiler, Developer Console, Scene Analysis, dashboard and release menus, with scripted substitutes.
• references/sources.md: every doc and video, with credentials and timestamps.
• scripts/rbx_perf.py: kit runner, offline audit, lab attributes, Server and Client windows, compare with a noise floor, profiler wrappers.
• scripts/agentkit/performance/: PerfProbe, ScriptProf, MicroProf, HeapProf, SceneScan, PerfAudit, lune_perf_audit.luau.