depths 0.2.0

S3-native OpenTelemetry compatible observability layer

Depths

Everything you need to build your observability stack — unified, OTel-compatible, S3-native telemetry. Built by Depths AI.

Depths covers the entire journey from ingestion of telemetry signals to persistence on S3 for cost optimization, and an efficient querying plan, including stats rollup to get snappy dashboards without having to touch raw logs.

Docs live at https://docs.depthsai.com.

---

Why Depths

• OTel first – accept standard OTLP JSON today, add protobuf by installing an extra
• Delta Lake by default – predictable schema across six OTel tables
• S3 native – seal a past UTC day, upload, verify rowcounts, then clean local state
• Polars inside – fast typed DataFrames and LazyFrames for compact reads
• Real-time + rollups – in-memory tail (SSE) and generalized stats sidecars (categorical & numeric)
• Simple to start – depths init then depths start

---

Install

# core (JSON ingest)
pip install depths

# optional protobuf ingest (OTLP x-protobuf)
pip install "depths[proto]"

---

Quick start

1) Initialize an instance

# plain init
depths init

# or initialize with schema add-ons
depths init --addons http,genai,db

This lays out ./depths_data/default with configs, index, day staging, and a local stats area (created on first use). When you pass --addons, your choices are saved to configs/options.json and applied automatically at runtime.

2) Start the OTLP HTTP server

# foreground
depths start -F

# or background
depths start

By default the service listens on 0.0.0.0:4318 and picks up the default instance without re-running init, using the options saved earlier (including any add-ons).

Customize:

depths start -F -I default -H 0.0.0.0 -P 4318

The server exposes:

• OTLP ingest: POST /v1/traces, POST /v1/logs, POST /v1/metrics

• Health: GET /healthz

• Reads:

  • Raw: GET /api/spans, GET /api/logs, GET /api/metrics/points, GET /api/metrics/hist

  • Stats (v0.2.0):

    • Register/remove:

      • POST /api/stats/categorical/add
      • POST /api/stats/numeric/add
      • POST /api/stats/remove

    • Query:

      • GET /api/stats/categorical
      • GET /api/stats/numeric

  • Real-time: GET /rt/{signal} where {signal} is traces | logs | metrics

3) Point your SDK or Collector

Most OTLP HTTP exporters default to port 4318. Example cURL for JSON:

curl -X POST http://localhost:4318/v1/logs \
  -H 'content-type: application/json' \
  -d '{"resourceLogs":[{"resource":{"attributes":[{"key":"service.name","value":{"stringValue":"demo"}}]},"scopeLogs":[{"scope":{},"logRecords":[{"timeUnixNano":"1710000000000000000","body":{"stringValue":"hello depths"}}]}]}]}'

If you installed the protobuf extra, you can send application/x-protobuf too.

---

The six Delta tables (layout)

Depths writes one Delta table per OTel family under a day root:

<instance_root>/staging/days/<YYYY-MM-DD>/otel/
  spans/
  span_events/
  span_links/
  logs/
  metrics_points/
  metrics_hist/

Real-time stream (SSE)

Peek at the newest telemetry as it arrives (before persistence). This is a best-effort tail; some items may never persist.

# logs stream
curl -N 'http://localhost:4318/rt/logs?n=100&heartbeat_s=10'

---

Schema add-ons

Depths can promote custom attributes to first-class columns across the six tables.

from depths.core.logger import DepthsLogger
from depths.core.config import DepthsLoggerOptions
from depths.core.schema import SchemaDelta

# add a custom top-level column on logs (example)
custom = SchemaDelta(
    name="my_app",
    columns={"region": str, "tenant_id": str, "tags": ("list", str)}
)

opts = DepthsLoggerOptions(addons={"logs": [custom, "http", "genai"]})
lg = DepthsLogger(options=opts)

Any remaining gen_ai.*, http.*, rpc.*, db.*, or geo.* attributes not promoted are preserved in their *_attrs_json column.

---

Generalized Stats sidecars (v0.2.0)

What you get

Two local Delta tables under ./stats/ that you control (opt-in per column):

<instance_root>/stats/
  stats_categorical/   # histograms for string/categorical columns
  stats_numeric/       # measures for numeric columns

Both tables are partitioned by: project_id / window / otel_table / column and roll per UTC minute buckets. The window controls the roll-up granularity: choose any of 1m, 5m, 15m, 30m, 1h, 1d per column.

Categorical rows (per bucket) carry:

• categories: list[str]
• counts: list[int]

Numeric rows (per bucket) carry population measures:

• event_count, value_min, value_max, value_mean, value_std, value_sum

Add/remove tracking (code)

from depths.core.logger import DepthsLogger

lg = DepthsLogger()

# Start histograms for a string column (multiple windows at once)
lg.stats_add_category(
    project_id="demo",
    otel_table="logs",
    column="http_method",
    windows=["1m","1h"]
)

# Start numeric measures for an int/float column
lg.stats_add_numeric(
    project_id="demo",
    otel_table="metrics_points",
    column="value",             # any numeric top-level column
    windows=["5m","1h","1d"]
)

# Stop a single (project/table/column/window) task at the next UTC minute
lg.stats_remove(project_id="demo", otel_table="logs", column="http_method", window="1m")

Notes

• Each window per column is an independent task; activation/deactivation happens on the next UTC minute boundary.
• Categorical aggregation learns categories on the fly; to avoid memory blow-ups, new categories beyond max_categories (default 200) are silently ignored.
• Stats sidecar can be disabled via options; sensible defaults mean most users won’t need to tweak StatsConfig directly.

Querying stats (HTTP)

# categorical: latest buckets for a column/window
curl 'http://localhost:4318/api/stats/categorical?project_id=demo&otel_table=logs&column=http_method&window=1h&latest_only=true&select=minute_ts,categories,counts'

# numeric: time-range query (epoch ms)
curl 'http://localhost:4318/api/stats/numeric?project_id=demo&otel_table=metrics_points&column=value&window=5m&start_ms=1710000000000&end_ms=1710086400000&select=minute_ts,event_count,value_mean,value_std'

Querying stats (Python)

# dicts (default)
rows = lg.read_categorical_stats(project_id="demo", otel_table="logs", column="http_method", window="1h", max_rows=100)

# Polars DataFrame
df = lg.read_numeric_stats(project_id="demo", otel_table="metrics_points", column="value", window="5m",
                           select=["minute_ts","event_count","value_mean","value_std"], return_as="dataframe")

---

Reading your data (raw tables)

Each endpoint accepts useful filters and returns JSON rows.

# last 100 logs with severity >= 9 that contain "error"
curl 'http://localhost:4318/api/logs?severity_ge=9&body_like=error&max_rows=100'

# metric points for a gauge/sum instrument
curl 'http://localhost:4318/api/metrics/points?project_id=demo&instrument_name=req_latency_ms&max_rows=100'

Programmatic reads:

from depths.core.logger import DepthsLogger

logger = DepthsLogger()
rows = logger.read_logs(body_like="timeout", max_rows=50)
print(rows[:3])

---

Identity context (opt-in)

Depths can enrich rows with session and user identity, following current OpenTelemetry attribute conventions. It’s off by default.

Enable via options (Python) or by editing configs/options.json:

from depths.core.logger import DepthsLogger
from depths.core.config import DepthsLoggerOptions

opts = DepthsLoggerOptions(
    add_session_context=True,
    add_user_context=True,
)

lg = DepthsLogger(options=opts)

When enabled, Depths reads these keys from event attributes first (then resource attributes):

• session.id → session_id
• session.previous_id → session_previous_id
• user.id → user_id
• user.name → user_name
• user.roles (list of strings) → user_roles_json (JSON-encoded)

When disabled, the columns remain empty.

---

S3 shipping

Turn on shipping and the background worker will seal completed days and upload them to S3, then verify remote rowcounts and clean the local day on a match.

S3 is configured from environment variables. A typical flow is:

1. Run with S3 configured in the environment
2. Depths rolls over at UTC midnight and enqueues yesterday for shipping
3. Shipper seals each Delta table, uploads, verifies, and cleans the local day

---

Configuration

• Instance identity and data dir come from DEPTHS_INSTANCE_ID and DEPTHS_INSTANCE_DIR (the CLI sets these).
• S3 configuration is read from environment variables.
• Runtime knobs (queues, flush triggers, shipper timeouts, stats cadence, real-time caps, identity context, and schema add-ons) live in the options object (depths.core.config.DepthsLoggerOptions). Add-on names or custom deltas are stored under addons in configs/options.json.

---

Development notes

• Package import is depths and can be installed with the protobuf extra using depths[proto].
• The service lives at depths.cli.app:app for uvicorn.
• CLI commands are available as depths init, depths start, and depths stop.

---

Status

Version v0.2.0. Adds generalized stats sidecars (categorical histograms and numeric measures with minute buckets, partitions by project/window/table/column, and developer-chosen windows 1m…1d), custom schema add-ons at init or via code, while keeping the real-time stream and the small read API.