latency-distribution-analyzer

Julia · HTTP.jl · Distributions.jl · StatsBase
Platform Reliability | SRE | Observability

Statistical latency analysis engine for production services. Ingests raw latency logs, fits optimal probability distributions via MLE, computes P50–P99.9 with bootstrap confidence intervals, and forecasts 24-hour SLA breach probability using a Markov chain model.

---

Why Julia?

Julia provides C-level numerical performance with Python-level ergonomics. For large-scale latency log processing, MLE distribution fitting, and Monte Carlo percentile bootstrapping, it outperforms Python/Pandas by 10–100× on hot numerical paths — without JVM overhead.

---

Features

• Distribution fitting — Evaluates Normal, LogNormal, Gamma, Weibull, Exponential; selects best by AIC
• Percentile bands — P50, P75, P90, P95, P99, P99.9 with 95% bootstrap confidence intervals
• SLA breach forecasting — Two-state Markov chain projects 24h breach probability
• REST API — POST /analyze, GET /health via HTTP.jl
• CLI mode — julia src/main.jl analyze <file.csv> <service> [sla_ms]
• Dockerized — single docker run to serve

---

Quickstart

Docker

docker build -t latency-analyzer .
docker run -p 8080:8080 latency-analyzer

Local

julia --project=. -e 'using Pkg; Pkg.instantiate()'
julia src/main.jl serve

CLI one-shot analysis

julia src/main.jl analyze data/sample_latency.csv checkout 200

---

API

POST /analyze

{
  "service": "checkout",
  "window_minutes": 60,
  "sla_ms": 200
}

Response:

{
  "service": "checkout",
  "sample_count": 1200,
  "sla_threshold_ms": 200,
  "best_fit_distribution": {
    "name": "LogNormal",
    "mean": 84.3,
    "std": 42.1,
    "params": "(4.28, 0.47)"
  },
  "percentiles": {
    "p500":  { "value": 72.4,  "ci_lo": 70.1,  "ci_hi": 74.8  },
    "p990":  { "value": 198.3, "ci_lo": 185.0, "ci_hi": 212.7 },
    "p999":  { "value": 341.2, "ci_lo": 310.4, "ci_hi": 378.9 }
  },
  "breach_probability_24h": 0.0312,
  "current_breach_rate": 0.028,
  "health": "green"
}

GET /health

{ "status": "ok", "version": "1.0.0" }

---

Architecture

CSV / Log Input
      │
      ▼
  Ingestion.jl  ──→  filter by service + time window
      │
      ▼
  DistributionFitter.jl  ──→  MLE fit (5 candidate distributions, AIC selection)
      │
      ├──→  Percentiles.jl  ──→  P50–P99.9 + bootstrap CI
      │
      └──→  SLABreach.jl    ──→  Markov chain 24h breach probability
                │
                ▼
           Report.jl  ──→  JSON + Markdown
                │
                ▼
           Server.jl  ──→  HTTP REST surface

---

Input Format

data/sample_latency.csv:

timestamp,service,latency_ms
2026-05-01T00:00:01,checkout,45.2
2026-05-01T00:00:02,checkout,52.7

Column	Type	Description
timestamp	ISO 8601	Request timestamp
service	String	Service identifier
latency_ms	Float	End-to-end latency in milliseconds

---

Environment Variables

Variable	Default	Description
PORT	8080	HTTP listen port
SLA_MS	200	SLA threshold in milliseconds
LATENCY_DATA	data/sample_latency.csv	Path to latency log CSV

---

Testing

julia --project=. test/test_distributions.jl

---

Stack

Package	Purpose
Distributions.jl	MLE distribution fitting
StatsBase.jl	Weighted percentiles, resampling
HTTP.jl	Lightweight REST server
CSV.jl + DataFrames.jl	Log ingestion
JSON3.jl	Request/response serialization
Optim.jl	Numerical optimization for MLE

---

Related Projects

Repo	Relationship
latency-budget-enforcer	Upstream: latency budget policy enforcement (Go)
agent-canary	Sibling: progressive rollout driven by latency signals
kinetic-flightdeck	Consumer: operator surface for platform health

---

License

AGPL-3.0 © Miz Causevic