pulselog 0.1.3

Real-time browser dashboard for Python logging — zero config, non-blocking

pulselog

Non-blocking Python logger with a live browser dashboard.

pip install pulselog

Every log.info() call costs 2.7µs. Zero config. Browser opens automatically.

---

Benchmark

Tested on Python 3.12.9, Windows, dashboard disabled (dashboard=False).

Scenario	Throughput	Notes
Single-thread burst	262.8k / sec	p50=2.7µs · p99=7.9µs · p99.9=87.9µs
Sustained (5s)	290.4k / sec	1.45M records logged
Queue saturation	318.1k / sec	0 drops — worker drained fast enough
Realistic (info + save + warn + error)	160.1k / sec	Mixed call types with kwargs
Multi-thread (8 threads)	41.3k / sec	0 drops — GIL is the ceiling, not PulseLog

A typical ML training loop logs 10–100 records/sec.
PulseLog handles 413× that load before any issues.

v0.1.1 → v0.1.2 improvements
  Single-thread   207k → 263k / sec    +27%
  p99 latency     11.9 → 7.9 µs        −34%
  Dropped records  16,011 → 0           FIXED

---

Quick start

from pulselog import Logger

log = Logger("my-app")

log.info("training started", epoch=1)
log.warning("learning rate too high", lr=0.1)
log.save("epoch-1", {"acc": 0.91, "loss": 0.23}, status="DONE", progress=33)

log.shutdown()

A browser tab opens at http://localhost:5678 and streams every log in real time.

---

Why pulselog?

Standard logging blocks the calling thread on every write — waiting for a file, a socket, or a database. In tight loops (ML training, data pipelines, inference servers) this adds up fast.

pulselog never blocks. Every log call enqueues a record in O(1) and returns immediately. A daemon worker drains the queue every 10ms and pushes batches to the dashboard over WebSocket.

log.info()         ← O(1), ~2.7µs, never blocks
      │
      ▼
  LogQueue          ← deque(maxlen=100_000), thread-safe, lock-free wake
      │
      ▼  every 10ms (adaptive — halves under load)
BackgroundWorker    ← daemon thread
      │
      ├──▶ DashboardServer.broadcast()  ← WebSocket → live browser
      │
      └──▶ (custom handlers)

---

Dashboard

Single self-contained HTML file served over WebSocket — no build step, no CDN, no npm.

Logs tab

• Colour-coded by level — DEBUG gray · INFO blue · WARNING amber · ERROR/CRITICAL red
• Level filter + full-text search
• Virtual list rendering — 100k+ logs with zero browser lag
• Auto-scroll with manual scroll override
• Export session as JSON

Checkpoints tab

• Progress bar per checkpoint
• Overall progress = average across all checkpoints
• Expandable JSON data viewer
• Status badges — DONE ✅ · IN_PROGRESS 🟡 · FAILED 🔴 · SKIPPED ⚫

---

API

Logger

log = Logger(
    name             = "my-app",
    host             = "localhost",
    port             = 5678,          # auto-increments if taken
    auto_open        = True,           # open browser on start
    dashboard        = True,           # False for CI / production
    checkpoint_path  = ".pulselog/checkpoints.db",
    level            = "DEBUG",
    worker_interval  = 0.01,           # drain interval in seconds (default 10ms)
    queue_size       = 100_000,        # max records before oldest evicted
    overflow         = "drop",         # "drop" | "block" | "raise"
)

Logging

log.debug("msg", **extra)
log.info("msg", **extra)
log.warning("msg", **extra)
log.error("msg", **extra)
log.critical("msg", **extra)

# kwargs appear as structured metadata in the dashboard
log.info("request handled", user_id=42, latency_ms=12, status=200)

# exception() captures the current traceback automatically
try:
    result = model.predict(x)
except Exception:
    log.exception("prediction failed", input_shape=str(x.shape))

Checkpoints

log.save(
    name      = "epoch-5",
    data      = {"loss": 0.31, "acc": 0.94},
    status    = "DONE",        # "DONE" | "IN_PROGRESS" | "FAILED" | "SKIPPED"
    note      = "best so far",
    progress  = 50             # 0–100, shown as progress bar in dashboard
)

result = log.load("epoch-5")        # → dict | None  (never raises)
names  = log.checkpoints()          # → list[str], most recent first
log.delete_checkpoint("epoch-3")

Context and grouping

# Tag groups subsequent logs under a label (per-thread — safe for concurrent use)
log.tag("training")

# Context manager — restores the previous tag on exit, even on exception
with log.context(tag="validation"):
    log.info("val loss", loss=0.41)
# tag is restored here

# Visual divider in the dashboard stream
log.divider("epoch boundary")

Utilities

log.flush(timeout=2.0)   # drain queue synchronously — returns False if timeout hit
log.shutdown()            # graceful teardown (also called automatically on exit)

stats = log.stats()
# {
#   "records_logged":   int,
#   "records_dropped":  int,
#   "drop_rate":        float,   # e.g. 0.04 = 4%
#   "queue_size":       int,
#   "queue_capacity":   int,
#   "queue_fill_pct":   float,
#   "checkpoints_saved": int,
#   "dashboard_clients": int,
#   "uptime_seconds":   float,
# }

---

stdlib logging integration

Drop-in bridge — all structured fields (lineno, filename, funcName, exc_info) are forwarded to the dashboard.

import logging
from pulselog.handler import PulseHandler

logging.getLogger().addHandler(PulseHandler("my-app"))

logging.info("this appears in the dashboard")
logging.error("with traceback", exc_info=True)  # traceback preserved

---

Configuration

Priority (highest → lowest): Logger() kwargs → env vars → pulselog.toml → defaults

Environment variables

PULSELOG_DASHBOARD=false
PULSELOG_HOST=0.0.0.0
PULSELOG_PORT=8080
PULSELOG_AUTO_OPEN=false
PULSELOG_CHECKPOINT_PATH=/data/checkpoints.db
PULSELOG_LEVEL=INFO
PULSELOG_WORKER_INTERVAL=0.01

pulselog.toml (place in project root)

[pulselog]
host            = "0.0.0.0"
port            = 8080
auto_open       = false
level           = "INFO"
worker_interval = 0.01

---

Production usage

# Disable the dashboard, keep checkpoints, log to stderr on drop
log = Logger(
    "prod",
    dashboard        = False,
    checkpoint_path  = "/data/checkpoints.db",
    overflow         = "drop",   # never block — warn on stderr instead
)

With dashboard=False:

• No threads started, no port bound, no browser opened
• Checkpoint reads/writes still work
• Log calls return in ~100ns (level check only, no queue overhead)
• CI environments (CI=true) disable the dashboard automatically

---

ML training example

from pulselog import Logger

log = Logger("resnet-training")

for epoch in range(1, 11):
    loss, acc = train_epoch(epoch)

    log.info(f"epoch {epoch}", loss=loss, acc=acc)

    log.save(
        f"epoch-{epoch}",
        {"loss": loss, "acc": acc},
        status   = "DONE",
        progress = epoch * 10,
    )

    if loss > prev_loss * 1.5:
        log.warning("loss spike detected", epoch=epoch, loss=loss)

log.shutdown()

---

Design notes

Queue — collections.deque(maxlen=100_000), guarded by a single threading.Lock. The _event.set() wake signal fires only on empty→non-empty transitions, not on every put(), reducing lock acquisitions under burst load.

Worker — wakes immediately on new records via threading.Event, falls back to polling every 10ms. Adaptive: halves the interval when queue exceeds 50% capacity, restores it when calm.

Drop policy — when the queue is full, the oldest record is evicted and a stderr warning is emitted every 1,000 drops. Configure overflow="block" to pause the caller instead, or overflow="raise" to surface the error explicitly.

Shutdown — atexit and SIGTERM both call shutdown() once (guarded against double-invocation). flush() accepts a configurable timeout and returns False if the queue wasn't fully drained in time.

Thread safety — tag() and context() use threading.local() so each thread maintains its own tag state independently.

---

Requirements

• Python ≥ 3.8
• websockets ≥ 11.0 (only needed with dashboard=True)

pip install pulselog            # includes websockets

---

License

MIT