SketchLog

Bounded-memory telemetry sketches for streaming metrics, live analytics, and production operations.

SketchLog compresses high-throughput telemetry streams into bounded-memory, mergeable statistical summaries. It combines DDSketch for latency percentiles, HyperLogLog for cardinality, and Count-Min Sketch for event frequencies, then adds production surfaces for APIs, dashboards, alerts, exporters, Kubernetes, multi-language SDKs, PostgreSQL-backed durability, and optional OmniKV embedded storage.

The goal is simple: keep the operational signal without retaining every raw event.

Why teams may use SketchLog

SketchLog is useful when you want fast operational answers from compact telemetry summaries instead of storing every raw event forever.

• Lower storage pressure: keep percentiles, cardinality, frequency, SLO, anomaly, and canary signals in bounded-memory summaries.
• Fast tail-latency answers: query p50, p95, and p99 without scanning large raw history.
• Streaming analytics: combine sketches with Streaming SQL, dashboards, anomaly comparison, Smart SLO workflows, and exporter payloads.
• Distributed and edge-friendly state: merge compact sketch state across nodes, tenants, and namespaces.
• Flexible durability: run ephemeral in memory, durable with PostgreSQL for server deployments, or embedded with OmniKV for local-first, edge, or single-node deployments.
• Proof-first evaluation: use the hosted playground, Docker smoke verifier, PostgreSQL durability proof, OmniKV storage proof, telemetry load proof, and public CI gates before trusting it with real workloads.
• Fits beside existing observability stacks: complement Prometheus, Mimir, Thanos, VictoriaMetrics, InfluxDB, TimescaleDB, Grafana, OpenTelemetry, Loki, Datadog, and New Relic instead of pretending to replace all of them.

SketchLog is currently a production-minded open-source beta. The core data structures and release pipeline are heavily tested, while some higher-level operational features continue to mature.

Live demo

Open the hosted playground: https://sbalavignesh123.github.io/sketchlog/demo/

The hosted playground is now a product evaluation hub: guided tour, synthetic dashboard, DDSketch percentiles, cardinality/frequency concepts, streaming SQL examples, stream operations, exporter payload previews, and proof commands for Docker, PostgreSQL, and OmniKV-backed storage paths.

Run the deterministic product demo and its end-to-end verifier:

docker compose -f demo/compose.yml up --build --wait

Open http://localhost:4173.

The demo includes live ingestion, percentile sketches, cardinality metrics, Streaming SQL examples, anomaly evidence, tenant isolation, exporter examples, and operational readiness checks. See the demo runbook for ports, cleanup, verification, and recording guidance.

Core capabilities

Architecture

flowchart LR
    subgraph Producers
        Python[Python SDK]
        TypeScript[TypeScript SDK]
        Go[Go SDK]
        OTel[OpenTelemetry]
        EBPF[Linux eBPF]
    end

    Python --> API[SketchLog API]
    TypeScript --> API
    Go --> API
    OTel --> API
    EBPF --> API

    API --> Registry[Namespace registry]
    Registry --> Stream[Bounded StreamLog]
    Stream --> DDS[DDSketch]
    Stream --> HLL[HyperLogLog]
    Stream --> CMS[Count-Min Sketch]
    Stream --> Analytics[SQL, SLO, diff, anomaly]
    Stream <--> Mesh[Authenticated Sketch Mesh]
    API --> Dashboard[Live dashboards]
    API --> Exporters[Prometheus, OTel, Loki, Datadog, New Relic]

Each stream stores compact sketch state rather than raw telemetry. See the architecture guide for merge behavior, windowing, drift detection, memory limits, and runtime details.

Installation

Install the embedded Python library:

pip install sketchlog

Install the standalone server:

pip install "sketchlog[server]"
sketchlog-server --host 127.0.0.1 --port 8000

Install the TypeScript client:

npm install @sketchlog/client

Install the Go client:

go get github.com/SBALAVIGNESH123/sketchlog/clients/go@v1.2.5

Quickstart

Python embedded sketch

from sketchlog import StreamLog

log = StreamLog()
log.add_latency(42.5)
log.add_batch([15.0, 88.2, 42.1])
log.add_unique("user_12345")
log.add_event("cache_miss", count=5)

print(f"p99 latency: {log.p99():.2f} ms")

Python async client

from sketchlog.async_client import AsyncSketchLogClient

async with AsyncSketchLogClient("http://localhost:8000") as client:
    await client.ingest_events(
        namespace="production",
        stream="api.latency",
        latencies=[42.5, 15.0, 88.2],
        uniques=["user_12345"],
        events={"cache_miss": 5},
    )

TypeScript client

import { SketchLogClient } from '@sketchlog/client';

const client = new SketchLogClient({ endpoint: 'http://localhost:8000' });

await client.ingestEvents('production_api', {
  latencies: [42.5, 15.0, 88.2, 42.1],
  uniques: ['user_12345'],
  events: { cache_miss: 5 },
});

Go client

import "github.com/SBALAVIGNESH123/sketchlog/clients/go"

client := sketchlog.NewClient(sketchlog.ClientOptions{
    Endpoint: "http://localhost:8000",
})

batch := sketchlog.EventBatch{
    Latencies: []float64{42.5, 15.0, 88.2, 42.1},
    Uniques:   []string{"user_12345"},
    Events:    map[string]int64{"cache_miss": 5},
}

err := client.IngestEvents(ctx, "production_api", batch)

Deployment

Run the server with Docker:

docker run --rm -p 8000:8000 ghcr.io/sbalavignesh123/sketchlog:1.2.5

Install with Helm:

helm upgrade --install sketchlog oci://ghcr.io/sbalavignesh123/charts/sketchlog \
  --version 1.2.5

Run mesh mode on Kubernetes:

helm upgrade --install sketchlog ./charts/sketchlog \
  --set replicaCount=3 \
  --set mesh.enabled=true \
  --set-string mesh.clusterSecret="$CLUSTER_SECRET"

SketchLog also includes Kubernetes Operator manifests and documentation for declarative cluster management.

Storage backends and proof commands

SketchLog stores compact stream summaries, not raw event rows. Choose the backend based on how you run SketchLog:

See the storage backends and proof guide for setup steps, tradeoffs, and reproducible evidence.

OmniKV is opt-in. To persist compact stream checkpoints and mesh tombstones into an embedded OmniKV database, install the OmniKV Python/native bridge and run:

export SKETCHLOG_STORAGE_BACKEND=omnikv
export SKETCHLOG_OMNIKV_DATA_DIR=/var/lib/sketchlog/omnikv
export SKETCHLOG_OMNIKV_NAMESPACE=sketchlog

sketchlog-server --host 0.0.0.0 --port 8000

This is opt-in. Existing in-memory and SQLAlchemy storage paths remain unchanged. See the OmniKV storage backend guide for the bridge contract and operational notes.

Run the unified storage proof CLI when you want one reproducible command for screenshots, launch videos, or local confidence checks:

python scripts/storage_proof.py --backend memory
python scripts/storage_proof.py --backend omnikv
python scripts/storage_proof.py --backend postgres --postgres-start --postgres-stop

The runner emits a human-readable summary and a JSON evidence report covering write, query, restart/reopen, delete, and durable tombstone behavior where the selected backend supports it.

Run the realistic telemetry load proof when you want evidence with real-ish API traffic instead of a tiny toy stream:

python scripts/telemetry_load_proof.py --backend memory
python scripts/telemetry_load_proof.py --backend omnikv
python scripts/telemetry_load_proof.py --backend postgres --postgres-start --postgres-stop

This proof deterministically generates JSONL-style production telemetry with timestamps, services, routes, statuses, users, tenants, regions, labels, and heavy-tailed latencies. It ingests the fixture through the HTTP API, verifies p50/p95/p99 through Streaming SQL, checks cardinality and top event counters, compares compact sketch memory against raw JSONL bytes, and verifies restart behavior for durable backends.

Documentation

The full documentation is published at sbalavignesh123.github.io/sketchlog/docs.

Important sections:

• Architecture
• Benchmarks
• Formal guarantees
• Client SDKs
• Async Python client
• Storage backends and proofs
• OmniKV storage backend
• Export integrations
• Kubernetes Operator
• RBAC
• Runbooks
• Threat model

Build docs locally:

pip install ".[docs]"
mkdocs serve

Development

Install the development environment:

make dev-install

Run checks:

make test
make test-go
make test-ts
make docs

Run the full demo:

make demo

What SketchLog is not

SketchLog is a bounded-memory telemetry analytics layer. It is deliberately not:

• A tracing system. It does not store request paths, spans, correlation IDs, or causal chains.
• A full time-series database. It does not store every raw sample, provide full historical drill-down, or implement complete label indexing like Prometheus, Mimir, Thanos, VictoriaMetrics, InfluxDB, or TimescaleDB.
• A raw log storage system. It keeps compact summaries, not every event payload.
• An exact analytics engine. Results are probabilistic with documented error bounds.

The intended positioning is complementary: use a TSDB when you need full raw history and label indexing; use SketchLog when you want compact streaming summaries, bounded-memory analytics, and proof-friendly operational signals.

Community

• GitHub: https://github.com/SBALAVIGNESH123/sketchlog
• Documentation: https://sbalavignesh123.github.io/sketchlog/docs/
• Slack: https://join.slack.com/t/sketchlog/shared_invite/zt-41kc03dnl-tiyHm4Gr2CbaJWuGHxdbiQ

License

SketchLog is released under the MIT License.