Skip to content

IMPROVEMENT_IDEAS.md

Latest commit

 

History

History
178 lines (143 loc) · 9.89 KB

IMPROVEMENT_IDEAS.md

File metadata and controls

178 lines (143 loc) · 9.89 KB
Raw

Monitoring Improvement Ideas & Speed Pattern Definitions

This document captures ideas for extracting richer telemetry (with or without a transparent proxy) and the current speed pattern heuristics implemented / proposed.

1. Connection & DNS

  • Fine‑grained phase timings via httptrace: DNSStart/Done, ConnectStart/Done, TLSHandshakeStart/Done, GotFirstResponseByte.
  • Record DNS server IP, answer count, A vs AAAA, per‑RR latency, cache indicator.
  • Resolve all returned IPs; optionally test each (multi-homing/CDN variance); annotate chosen vs fastest.
  • TCP RTT estimate (SYN timing) or light ICMP/TCP ping before transfer; store min/avg.

2. TLS / Security

  • TLS version, cipher suite, ALPN (h2 vs http/1.1), session resumption reused?, key exchange type.
  • Certificate: issuer CN, SAN count, days to expiry, OCSP stapling presence.
  • MITM / proxy clues: unexpected issuer, downgraded TLS version, missing ALPN, added headers (Via, X-Forwarded-For).

3. HTTP Layer

  • Status code & redirect chain with per-hop timing.
  • Key headers: Server, Via, Age, Cache-Control, Content-Length, Content-Encoding, ETag, Last-Modified.
  • Separate metrics: TTFB vs transfer duration.
  • Compression ratio: uncompressed_estimate / wire_size if Content-Encoding present.
  • Chunked vs fixed length detection.

4. Throughput & Stability Metrics

  • Existing: average, stddev, pattern detection.
  • Add: median, min, max, p50, p90, p95, p99, IQR, coefficient of variation.
    • p50 (median): typical sustained speed; robust against outliers; good baseline.
    • p90: 90% of samples are at or below this; gap between p50 and p90 shows higher-end burst capacity.
    • p95: highlights occasional higher spikes; large p95 - p50 indicates intermittent bursts (maybe congestion relief moments).
    • p99: tail of distribution; if p99 >> p95 the very top speeds are rare spikes (bufferbloat recovery, transient caching, TCP window growth spurts).
    • Interpretation patterns:
      • Narrow spread (p99/p50 < 1.2) => very stable link.
      • Moderate spread (1.2–1.8) => normal WAN variability.
      • High spread (>1.8) with high stddev => volatile; possible congestion, policing, or shared medium contention.
      • If slope positive but p90 ~ p50 (little headroom) => slow ramp but quickly plateaus (maybe high RTT / small initcwnd).
      • High p99 with low median => short opportunistic bursts (possibly queue draining or short-lived cross-traffic gaps).
  • Rolling jitter: mean absolute percent change between consecutive samples.
  • Plateau detection: consecutive samples within ±X% threshold (count & durations).
  • Linear regression slope on (time, speed) for ramp/decline classification.
  • Goodput in first RTT window (approx initial congestion window behavior).

5. Advanced Transfer Scenarios

  • Two-phase test: small object (latency) + large object (throughput plateau).
  • Parallel fetches (N identical downloads) for fairness / per-flow throttling detection.
  • Range requests: verify Accept-Ranges; compare partial vs full speed.
  • Warm cache test: immediate repeat request diff.

6. Proxy / Caching Detection

  • Compare DNS resolved IP(s) vs conn.RemoteAddr(); mismatch suggests interception.
  • Headers: Via, X-Cache, Age; detect cache hit (Age > 0 or X-Cache: HIT).
  • Inject custom header (e.g., X-Probe: <rand>); confirm unmodified echo/absence.
  • Timing difference HEAD vs GET; large discrepancy may indicate proxy prefetching.

7. Path & Geography

  • GeoIP ASN + ISP/Org (ASN DB).
  • Infer hop count from initial TTL (64/128/255 heuristics) minus observed TTL if captured (requires raw socket or pcap/traceroute).
  • Optional asynchronous traceroute; store path signature hash to detect changes.

8. Environment & Host Context

  • Local interface, IP, MTU.
  • Active congestion control algorithm (/proc/sys/net/ipv4/tcp_congestion_control).
  • System load averages, CPU steal %, network namespace identifier.
  • Tool version + build hash for reproducibility.

9. Robustness & Error Handling

  • Unified error taxonomy: dns_error, tcp_error, tls_error, http_error, read_error.
  • Retry policy with backoff for transient categories; record attempt count.
  • Capture partial metrics even on failure (e.g., DNS + TCP times when TLS fails).

10. Data Model Enhancements

  • schema_version field.
  • Embed configuration snapshot (sample_interval_ms, parallel_downloads, user_agent, extended_tests_enabled).
  • Executable hash (SHA256), build timestamp.

11. Anomaly Flags (Booleans)

  • possible_proxy
  • cache_present
  • volatility_high
  • ramp_slow
  • decline_detected
  • mid_dip
  • plateau_stable

12. Optional Stretch Goals

  • QUIC/HTTP3 probe for latency/throughput comparison.
  • Passive loss inference: detect long read stalls > threshold.
  • DNS over HTTPS vs classic DNS comparative timing.

Current / Proposed Speed Pattern Heuristics

(Implemented or suggested in monitor.go)

Pattern Key (description) Trigger Logic (simplified)
slow start until max speed first < 0.5 * max
good speed at begin but slow in the end last < 0.5 * max
highly volatile speed measurement over whole window stddev > 0.5 * mean
stable speed throughout transfer stddev < 0.1 * mean
fast start, reaches max speed quickly first > 0.8 * max
fast end, maintains speed until finish last > 0.8 * max
mid-transfer dip: speed drops in the middle mid < 0.5 * max & first,last > 0.7 * max
slow start, then speed recovers mid-transfer first < 0.5 * max & mid > 0.8 * max
gradual decline: speed decreases throughout monotonic decreasing & first > last
gradual increase: speed increases throughout monotonic increasing & last > first

Potential Additional Patterns

  • Oscillating: alternating up/down beyond a % threshold.
  • Early plateau: reaches ≥0.9 * max within first 20% time and stays ±10%.
  • Late surge: last 20% includes new max > 1.1 * previous max.
  • Multi-phase: two distinct plateaus separated by dip > X%.

Implementation Ordering Suggestion

  1. Add config + httptrace instrumentation.
  2. Extend schema with version + config snapshot + basic TLS/meta.
  3. Add percentile & regression-based stability stats.
  4. Introduce anomaly flags & refined pattern detection.
  5. Add proxy/cache detection logic (headers, repeat request).
  6. Add multi-IP and ASN/Org enrichment.
  7. Parallel / range / warm cache tests.
  8. (Optional) Traceroute + QUIC.

Notes

  • Keep additional tests optional (config flags) to control runtime cost.
  • Persist raw samples; derived metrics can be recomputed offline.
  • Document schema changes with version bumps.

TODO / Bugs

  • Viewer: Screen not updated when filtered batches == 1
    • Repro:
      1. Open viewer and load a results file with at least two situations, where one situation has only 1 batch.
      2. Select that situation from the Situation drop-down (Batch X-axis mode).
      3. Observe: table updates, but sometimes the chart image doesn’t change (appears stale) even though filteredSummaries returns 1.
    • Notes:
      • We already pad X-axis ranges and duplicate the single point for go-chart, and we refresh canvases/overlays.
      • Likely a repaint race or image cache edge when dimensions don’t change. Possible fix: force a two-step image swap (set Image=nil and Refresh, then set new image and Refresh) or swap to a 1px different MinSize before setting back.
      • Also check Time axis path for single timestamp and verify legend/series changes trigger a repaint.

Backlog / Sizing & Font Uniformity

  1. Chart Font Uniformity Test & Enforcement

    • Add a lightweight test in cmd/iqmviewer/main_test.go that invokes the custom renderers (detailed Host/IP and batch Host/IP timing charts) and inspects chosen font sizes (e.g., by wrapping loadDynamicFontFace or exposing selected size via a debug hook) to ensure they do not regress below a minimum readable threshold (e.g., 14pt for medium bars, 20pt for large bars).
    • Provide a central mapping for (barHeight -> fontSize) to avoid duplicate switch statements; simplifies future global adjustments.
    • Consider an environment var override (e.g., IQM_VIEWER_FONT_SCALE=1.2) for quick user scaling.

  2. User-Selectable Sizing Modes

    • Add a dropdown or settings dialog (Compact / Normal / Large / Huge) that influences bar height & gap scaling factors across all custom charts for consistent look.
    • Persist preference in existing viewer prefs.

  3. Scroll vs. Compress Strategy

    • For > 40 batches, optionally allow a vertical scroll container instead of shrinking bar heights; preserves legibility for large historical windows.
    • Add toggle: "Auto Compress" vs "Scroll".

  4. Legend / Phase Key Panel

    • Provide a small legend panel (DNS/TCP/TLS/Wait/Transfer/Stall colors) reused by both detailed and batch Host/IP charts; reduces need for repeated inline labels when bars become very wide.

  5. Export Font Consistency

    • Ensure export re-render path uses same sizing logic but can optionally upscale font slightly (e.g., +2pt) for high-DPI clarity.
    • Add integration test: export all charts to a temp dir and validate PNG dimensions & presence; future enhancement could parse PNG text chunk to confirm watermark.

  6. Performance Optimization (Single-Pass Aggregation)

    • Replace per-batch file rescans with a single streaming pass building structures for all batches, then compute both detailed and batch aggregate timing charts. Improves load time for large result sets.

  7. Accessibility / High Contrast Mode

    • Offer alt color palette for color-blind accessibility (e.g., ColorBrewer Set2 or Tol palette) selectable via settings.

  8. Optional Global Font Override

    • Config field to specify a TTF path; load once and propagate to all chart renderers for uniform typography.

  9. Unit Tests for Sizing Regression

    • Store representative synthetic batch counts (e.g., 4, 8, 16, 32, 40) and assert computed bar height & font size pairs against expected golden values.

  10. Watermark & Title Scaling

    • Scale watermark and title text proportionally to base font to avoid tiny watermark on very tall charts.