beava 0.0.4

Beava: real-time feature server with composable streaming pipelines (Python SDK + bundled Rust server binary).

beava — Python SDK

Python SDK for Beava, the single-binary real-time feature server. Author pipelines with @bv.event / @bv.table decorators; push events and read features against a running beava server over HTTP or TCP.

Install

# pip    (recommended) — installs SDK + bundled Rust server binary from PyPI
pip install beava

# brew   — Homebrew formula (macOS + Linuxbrew)
brew install beava-dev/beava/beava

# docker — zero deps on the host
docker run -p 8080:8080 -p 8081:8081 beavadev/beava:latest

The wheel ships the SDK and the Rust beava server binary (polars-style); after install, the beava shell command is on PATH and the SDK can run against it directly — including embed mode (bv.App() with no URL). Pin a specific version with pip install beava==0.0.0.

Quickstart

import beava as bv

@bv.event
class Click:
    user_id: str
    page: str

@bv.table(key="user_id")
def UserActivity(e: Click):
    return e.group_by("user_id").agg(
        clicks_1h=bv.count(window="1h"),
        unique_pages_1h=bv.n_unique("page", window="1h"),
    )

app = bv.App(url="http://localhost:8080")
app.register(Click, UserActivity)

app.push("Click", {"user_id": "alice", "page": "/home"})
app.push("Click", {"user_id": "alice", "page": "/products"})

app.get("UserActivity", "alice")
# => {"clicks_1h": 2, "unique_pages_1h": 2}

Transports

The same SDK speaks three transports — pick by the URL you pass to bv.App(...).

# HTTP/JSON (curl-compatible debugging path)
app = bv.App(url="http://localhost:8080")

# Framed TCP (sub-millisecond fast-path; JSON or msgpack content)
app = bv.App(url="tcp://localhost:8081")

# Embed mode (no separate server — auto-spawns a local `beava` binary)
app = bv.App()

Embed mode finds the beava binary that ships with the curl-installed wheel automatically. Override with BEAVA_BINARY=/path/to/beava if you want to point at a different build.

Surface

Method	Wire	Notes
app.register(*descriptors, force=False, dry_run=False)	POST /register	Returns {"registry_version": <n>}. Pass force=True for destructive schema changes; dry_run=True returns a categorized diff without applying.
app.push(event_name, fields)	POST /push body {event, data}	Fire-and-forget. Returns server ack.
app.get(table, key=None, features=None)	POST /get body {table, key}	Returns a flat dict of feature values. key=None routes to the global-aggregation sentinel (ADR-003). features=[...] narrows to a subset.
app.batch_get(requests)	POST /batch_get body {requests: [...]}	requests is a list of (table, key) or (table, key, features) tuples. Returns a list of dicts in input order.
app.reset()	POST /reset (test_mode-only)	Wipes server state. Pass test_mode=True to bv.App(...) to enable.
app.ping()	POST /ping	Liveness; returns {"pong": True, "registry_version": <n>}.
app.close()	—	Releases the transport connection.

Pipeline DSL

@bv.event declares an event schema. Fields are typed Python class attributes; the SDK serializes them to JSON on push.

@bv.table(key=...) declares a feature table. The decorated function takes one event-typed parameter and returns e.group_by(...).agg(...) over it. The runtime maintains the table incrementally — every event updates exactly the affected key's row.

The key= kwarg accepts a string (single-column key) or a list of strings (composite key).

bv.col("name") references an event field inside a where= filter or arithmetic expression. Strings in where= are rejected — pass an _Expr (e.g. bv.col("event_type") == "click").

bv.lit(value) wraps a Python literal so it can participate in the same expression grammar.

Operator catalogue

50+ aggregation primitives are exported flat at the bv.* namespace. Inspect the live surface from Python:

import beava as bv
print([x for x in dir(bv) if not x.startswith("_")])

Selection by family:

• Core — count, sum, mean, min, max, var, std, n_unique, quantile
• Sketch — top_k, bloom_member, entropy, histogram, hour_of_day_histogram, dow_hour_histogram
• Recency — first, last, first_n, last_n, lag, first_seen, last_seen, age, time_since
• Decay — ewma, ema, ewvar, ew_zscore, decayed_sum, decayed_count
• Velocity — rate_of_change, inter_arrival_stats, burst_count, delta_from_prev, trend, z_score
• Buffer / geo — most_recent_n, reservoir_sample, geo_velocity, geo_distance, geo_spread, distance_from_home

Deprecation aliases retained: avg → mean, variance → var, stddev → std, count_distinct → n_unique, percentile → quantile. Use the canonical form in new code.

Demo data

bv.demo exposes three pre-built pipelines + dataset loaders for trying the SDK without writing your own pipeline:

import beava as bv
adtech = bv.demo.adtech()      # ad-impression / click / conversion
fraud = bv.demo.fraud()        # transaction fraud
ecommerce = bv.demo.ecommerce()  # cart / view / purchase

Each returns a tuple of (events, tables, dataset) ready to register and push.

Errors

Top-level error types: bv.RegistrationError (raised on register failures with structured code + message), bv.ValidationError (raised by client-side input validators), bv.BinaryNotFoundError (raised in embed mode if the beava binary isn't on PATH).

Learn more

• beava.dev — site, guide, docs
• Root README — server install, wire surface, server CLI
• Discord — questions and feedback