Logging, signposts & power — from code to field data

Unified logging — how it works

• Logger(subsystem: "com.acme.app", category: "network") — subsystem = reverse-DNS app/module, category = component; both are filter keys.
• The message is stored binary: format string + arguments, rendered when someone reads it. Calling it is cheap; a disabled debug is almost free. print builds a string and only reaches stdout.
• Levels (Apple’s table): debug not persisted · info persisted only when collected with the log tool · notice (default), error, fault persisted up to a storage limit; error/fault also capture the activity’s process chain. Aliases: trace=debug, warning=error, critical=fault.
• Privacy: interpolated ints, floats, bools are public; dynamic strings and objects are redacted (<private> when no debugger is attached). Opt in per value: privacy: .public, or .private(mask: .hash) to correlate without revealing.
• Read: Xcode 15+ console is structured (level, subsystem, category, source line). Console.app: pick the device, Include Info/Debug Messages. Field bugs: sysdiagnose, or read your own logs in-app with OSLogStore(scope: .currentProcessIdentifier) and attach them to a report.

Example

import os
extension Logger {
  static let net = Logger(subsystem: "com.acme.app",
                          category: "network") }
Logger.net.info("GET \(url.path, privacy: .public)")
Logger.net.error("HTTP \(code) user \(uid,
                 privacy: .private(mask: .hash))")
let sp = OSSignposter(logger: .net)
func fetch() async throws -> Feed {
  let st = sp.beginInterval("Fetch", id: sp.makeSignpostID())
  defer { sp.endInterval("Fetch", st) }  // same StaticString name
  sp.emitEvent("CacheMiss")               // a point, not a span
  return try await api.feed()
}

log stream --level debug --predicate 'subsystem == "com.acme.app"'
xcrun xctrace record --template 'Time Profiler' --time-limit 30s \
  --output feed.trace --launch -- MyApp.app     # CI, no GUI

Signposts — time, not words

• Interval = begin/end with the same static name; the OSSignpostID tells overlapping intervals apart. Event = one instant. Category .pointsOfInterest lands in the Points of Interest track next to Time Profiler samples.
• Near-free when nobody is recording. Xcode 27: the os_signpost instrument shows a track per signpost name under its category.
• CI: XCTOSSignpostMetric(subsystem:category:name:) in measure(metrics:); xctrace: --run-name (Xcode 26), --recording-options <json> (27); xctrace export takes .atrc/.logarchive directly and a time range (27).

Energy — what costs power

Power Profiler (Xcode 26)

• Instrument on an iOS/iPadOS device: a system power lane (with thermal + charging state) and your process’s power impact per subsystem — CPU, GPU, display, networking. The score finds spikes; fix the highest-impact subsystem first.
• Two modes: tethered (Launch/Attach — needed for per-process metrics) or passive — start a trace in the device’s Developer settings, use the app normally, open the .atrc in Instruments (power + lower-rate CPU samples). The Energy Impact gauge in Xcode is the quick live view.

Field data — MetricKit & Organizer

• MetricKit: MXMetricManager.shared.add(subscriber); MXMetricPayload ≈ daily per device, …Metrics for: cpu, gpu, display, locationActivity, networkTransfer, cellularCondition, applicationLaunch, applicationResponsiveness (hangs), animation (hitches), memory, diskIO, applicationExit, signpost. You upload it.

• Your own field timings, aggregated into signpostMetrics:

  let h = MXMetricManager.makeLogHandle(category: "Feed")
  mxSignpost(.begin, log: h, name: "Load")  // ... .end

• Organizer: Apple aggregates opt-in users; needs volume, compares versions. Battery Usage, Launch Time, Hang Rate (s of hang per hour), Memory, Disk Writes, Terminations. Xcode 27: Hitches replaces Scrolling (all animations); Storage (Documents & Data, App Size); Insights Overview ranks regressions; metric goals (Battery, Disk Writes, Hang Rate, Hitches, Memory, Storage).

Interview traps

• print/NSLog for diagnostics: eager strings, no levels, no redaction; print never reaches a sysdiagnose.
• Support asks for logs and there is nothing: debug/info weren’t persisted — log what support needs at notice+.
• privacy: .public on a token or email — that is a PII leak.
• Energy measured in the Simulator, or plugged in and idle.
• Organizer is not real-time: aggregated, opt-in, needs users.

Remember

Logger = words · signposts = time · Power Profiler = joules · MetricKit / Organizer = the field.

Likely questions

1. Which levels persist? — notice/error/fault; info if collected; never debug.
2. Why Logger over print? — cheap, levelled, redacted, persisted.
3. Time a flow for real users? — mxSignpost → signpostMetrics.
4. Battery complaints? — Organizer + MetricKit, repro with Power Profiler.