# Performance and footprint

Know what the Screeb SDK adds to your app in size, memory, CPU, network and storage, with and without session replay.

The figures below come from a repeatable, versioned measurement on a release build. They cover two separate regimes:

- **Core**: surveys and in-app messages, what every integration gets.
- **Session replay**: screen recording, opt-in only. It has no runtime cost (CPU, memory, network) until it starts, from the SDK or through the targeting you set in Screeb. Its code ships in the SDK binary in every case.

Exact numbers vary by device and OS version.

Measured on Screeb iOS SDK 4.0.4 (2026-08-20), on a physical iPhone 17 Pro running iOS 26.6.

## Summary

| Metric     | Core (surveys, default)                                                                                                                | Session replay (opt-in)                                                                                                                        |
| ---------- | -------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
| App size   | About 0.33 MB compressed (about 1.07 MB uncompressed framework, arm64 device slice)                                                    | Included: same framework                                                                                                                       |
| Memory     | About 4 MB in your app's process. The rendering surface (a `WKWebView`) runs in iOS system WebKit processes, outside your app's memory | About 3 to 4 MB more while recording, with short peaks while a full frame is encoded                                                           |
| CPU        | About 0.1% of one core: the SDK is idle between events                                                                                 | Around 6 to 12% of one core on a continuously changing screen, far less on mostly static screens, zero when not recording                      |
| Frame rate | No measurable impact                                                                                                                   | The recorder's main-thread budget is capped and follows the display refresh rate. The app's median frame time stays at 16.7 ms while recording |
| Network    | About 2 MB on the first session for the survey bundle, then served from the HTTP cache. Event tracking is batched                      | Scales with how much the screen changes: up to about 1.5 MB per minute on continuously animated content, near zero on static screens           |
| Storage    | None managed by the SDK: no files, no databases, no `UserDefaults`. The WebView keeps a few KB of `localStorage` and its HTTP cache    | None beyond Core: frames are never written to disk                                                                                             |

## App size

The SDK adds about 0.33 MB compressed to an iOS app. The embedded framework is about 1.07 MB on disk for the arm64 device slice, before App Store thinning and compression. This figure covers Core and session replay, which ship in the same framework: there is no separate module. The exact App Store download size depends on app thinning for each device model.

## Memory

The SDK renders surveys and messages in a `WKWebView`. WebKit runs its content, networking and GPU work in separate system processes managed by iOS, not in your app's memory. In the reference measurement, the SDK added about 4 MB to the app process, measured as `phys_footprint`, the figure iOS uses to decide which apps to reclaim. Session replay adds about 3 to 4 MB more while recording (a capture bitmap and a tile cache), with short peaks while a full frame is encoded.

## CPU

At rest, the SDK adds no measurable CPU cost: it is event-driven and idle between interactions, and survey rendering happens in the WebKit processes.

Session replay uses CPU only while it records, and the cost follows how much the screen changes:

| Screen activity while recording                                       | CPU on the reference device                                                                                                                                                                   |
| --------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Continuously changing (permanent scrolling and full-screen animation) | About 6% of one core on a 120 Hz (ProMotion) display, where the recorder spaces captures to respect the 8.3 ms frame budget. Up to about 12% on 60 Hz hardware, which allows a faster cadence |
| Localized but permanent change (a ticking widget)                     | About 5 to 8% of one core                                                                                                                                                                     |
| Occasional interaction, static in between                             | About 2 to 3% of one core (heartbeat checks)                                                                                                                                                  |
| Not recording                                                         | Zero                                                                                                                                                                                          |

The recorder's main-thread budget is capped: capture frequency is bounded to a small share of your app's frame budget, and diffing, encoding and sending all run off the main thread. In the benchmark, the app's median frame time stays at 16.7 ms (60 fps) while recording; only the p95 moves, under continuous full-screen animation. Unchanged frames are skipped before any encoding, so a quiet screen costs almost nothing.

The recorder also adapts to the device. On ProMotion displays the frame budget is halved, so captures go in the idle tail of the frame and are spaced further apart. Under thermal pressure, Low Power Mode or memory pressure, it reduces its work: quality and resolution first, cadence last (see [battery usage](https://developers.screeb.app/sdk-ios/session-replay#battery-usage)). The figures above were measured on a cool device; a heat-soaked phone measured lower CPU still.

## Network

A first survey session downloads the survey rendering bundle, the versioned web runtime shared by the Screeb SDKs: about 2 MB, measured on Android where WebView traffic is attributable per app. It is served with standard HTTP caching, so it is not downloaded again at each session, only when the bundle version changes. Event tracking is batched and adds close to no traffic.

Session replay sends only the regions that changed, so its traffic follows how much the screen changes:

| Screen activity while recording     | Replay traffic on the reference device                                                                                          |
| ----------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
| Continuously changing               | About 0.9 MB per minute on the 120 Hz reference device, up to about 1.5 MB per minute where the display allows a faster cadence |
| Localized change (a ticking widget) | About 0.3 MB per minute                                                                                                         |
| Static between interactions         | About 0.05 MB per minute                                                                                                        |
| Not recording                       | None                                                                                                                            |

Treat the first row as an upper band for very active content, not a fixed rate: sessions with occasional interaction stay well below it.

## Storage

The SDK keeps no persistent storage of its own: no files, no databases, no `UserDefaults`. Its `WKWebView` keeps a few KB of session state in `localStorage` and its own HTTP cache for the survey bundle, both managed by WebKit. Session replay never writes frames to disk.
