eka-usage-sdk

Monorepo of language-specific SDKs that push metered usage and critical logs from Eka Care services to Kafka.

Install

Python (pip / uv):

pip install git+https://github.com/eka-care/eka-usage-sdk.git#subdirectory=sdks/python

Or in pyproject.toml:

dependencies = ["eka-usage-sdk @ git+https://github.com/eka-care/eka-usage-sdk.git#subdirectory=sdks/python"]

Go:

go get github.com/eka-care/eka-usage-sdk/sdks/go

Build with -tags confluent to link the real Kafka producer.

TypeScript / Node.js (published to GitHub Packages on ts-v* tags):

npm install @eka-care/usage-sdk

Configuration

Kafka connection and tuning come from environment variables — no hardcoded broker addresses in service code.

Env var	Default	Notes
EKA_KAFKA_BROKERS	—	Required
EKA_KAFKA_LINGER_MS	50	Python / Go only (librdkafka)
EKA_KAFKA_BATCH_SIZE	65536	Python / Go only (librdkafka)
EKA_KAFKA_COMPRESSION_TYPE	lz4	All SDKs
EKA_KAFKA_ACKS	1	All SDKs
EKA_KAFKA_RETRIES	5	All SDKs

Explicit constructor args always override env.

API surface

All three SDKs expose the same three-method surface. One client per process serves all tenants — workspace_id is passed per call.

EkaClient(service_name, ...)
client.record(workspace_id, product, metric_type, quantity=1.0, status="ok", unit_cost=None, metadata={})
client.shutdown()

When status="error", the metadata JSON string serves as the error log.

Products and metric types

Product	Metric types
ekascribe	transcription_minute, transcription_session
mr_ai	mr_record_upload, mr_page_processed
agent	chat_session, tool_call, tool_call_error, credit_consumed, input_token, output_token
api	api_call, api_error
webhooks	webhook_push, webhook_delivery_failed
emr_tools	tool_call, tool_call_error
clinical_tools	tool_call, tool_call_error
comms	sms, whatsapp, email
abdm	api_call, abha, consent, fetch, storage

Statuses: ok (billable, is_billable=1) or error (non-billable, is_billable=0)

The SDK validates these at call time — invalid values go to on_error, never to Kafka.

All events flow through a single Kafka topic:

Topic	Partitions	Retention	Key
eka.usage.events	10	7 d	workspace_id

A Kafka Connect ClickHouse sink (infra/kafka-connect-sink.json) consumes the topic downstream. The SDK has no knowledge of ClickHouse.

Principles

• Never block the caller. record returns immediately.
• Never crash the host. All Kafka errors are caught and routed to on_error.
• Never require connection management. The SDK owns the producer lifecycle.
• Always enrich server-side fields. ts, sdk_version, and hostname are set inside the SDK so callers cannot spoof them.
• Always validate before producing. Invalid events never reach Kafka — they go to on_error.

Layout

eka-usage-sdk/
├── README.md
├── Makefile                       # test-all / lint-all / build-all
├── .github/workflows/
│   └── publish-ts.yml             # publish TS SDK to GitHub Packages
├── infra/
│   ├── create-topics.sh           # one-shot Kafka topic setup
│   └── kafka-connect-sink.json    # ClickHouse sink config
├── sdks/
│   ├── python/
│   ├── typescript/
│   └── go/
└── docs/
    ├── integration-guide.md
    └── metric-types.md

Quick start

# Run every language's tests (uses mock Kafka producers)
make test-all

# Lint everything
make lint-all

# Build every artifact
make build-all

Per-SDK READMEs under sdks/<lang>/README.md cover language-specific install and usage. The canonical list of valid products, metric types, statuses, and log levels lives in docs/metric-types.md — changes there must be mirrored in all three SDK constants files.

Publishing

• Python / Go: Services install directly from this repo (no publish step).
• TypeScript: Push a tag like ts-v0.2.0 to trigger the GitHub Actions workflow that publishes to GitHub Packages. Bump version in sdks/typescript/package.json before tagging.