perf(client): exponential backoff in UploadRenderedProfile — ΔT≤5s vs const 5s sleep

[KorsarOfficial]

Closes #93

Problem

UploadRenderedProfile contained an unconditional time.Sleep(5s) with a // FIXME comment. ClickHouse async_insert makes the uploaded profile temporarily invisible, but the fixed sleep adds a constant 5-second penalty to every call, including the common case where data is available in < 100 ms.

Complexity / latency analysis

Let T_vis = actual ClickHouse visibility latency (random variable).

Strategy	E[latency]	Worst case
Sleep(5s)	5 s	5 s
Exponential backoff (this PR)	E[min(T_vis, 5s)] + O(log(5s/100ms)) overhead	5 s

Backoff sequence: 100 ms, 200 ms, 400 ms, 800 ms, 1600 ms, then capped at remaining.
Number of poll attempts: at most ⌈log₂(5000/100)⌉ + 1 = 7.

For T_vis ~ Uniform(0, 200 ms): E[latency] ≈ 100 ms vs 5000 ms — 50× improvement.

Implementation

const maxWait = 5 * time.Second
delay := 100 * time.Millisecond
deadline := time.Now().Add(maxWait)
for {
    taskID, _, err = c.MergeProfilesProto(ctx, req, taskAnnotation)
    if err == nil { break }
    remaining := time.Until(deadline)
    if remaining <= 0 { return "", "", fmt.Errorf("failed to render uploaded profile: %w", err) }
    if delay > remaining { delay = remaining }  // clamp to deadline
    select {
    case <-ctx.Done(): return "", "", ctx.Err()
    case <-time.After(delay):
    }
    delay *= 2
}

Invariants:

• Total wait ≤ maxWait = 5s (unchanged worst-case reliability)
• Context cancellation respected — no goroutine leak
• delay is clamped to remaining before the last sleep, preventing overshoot

[KorsarOfficial]

📄 Full analysis report (PDF): 08-perforator-optimizations.pdf

Covers complexity analysis, concurrency audit, and verification for all 7 optimizations.