What is an event analytics pipeline and what events does it collect?

An event analytics pipeline is a system that ingests, stores, and analyzes discrete user actions and system events to understand behavior patterns. It collects events such as page views, button clicks, form submissions, purchases, errors, feature usage, search queries, and session lifecycle events (app open, app close). Each event carries a schema with a name, timestamp, user/device identifiers, and a property bag of context-specific fields. The pipeline receives events from client SDKs (web, iOS, Android) and server-side instrumentation, writes them to a durable event store (Kafka, Kinesis), and routes them to an OLAP database for querying and to a stream processor for real-time aggregation.

How does an event analytics collection SDK handle offline buffering and reliability?

An event analytics SDK buffers events locally on the device when the network is unavailable, writing them to a persistent queue (SQLite, local storage, or a file) so they are not lost if the app crashes or the device goes offline. Events are batched by count (e.g., 50 events) or time interval (e.g., 30 seconds) before being sent to the ingestion endpoint to reduce the number of HTTP requests. On send failure the SDK applies exponential backoff with jitter and retries from the persistent queue. Each event is assigned a client-generated UUID so the server can deduplicate retried submissions. When the device comes back online the SDK drains the queue in order, ensuring all offline events are eventually delivered with correct timestamps.

How is event data stored in a columnar OLAP database for fast analytics?

Event data is stored in a columnar OLAP database such as ClickHouse, Apache Druid, or BigQuery by organizing each event property as a separate column stored contiguously on disk. Queries that aggregate a single property (e.g., COUNT(*) WHERE event_name = 'purchase') only read that column's data, skipping all other columns, which minimizes I/O. Columnar storage also achieves high compression ratios because repeated values in a column (e.g., the same event_name appearing millions of times) compress very efficiently with dictionary or run-length encoding. Events are partitioned by date and sorted by (event_name, user_id, timestamp) so range scans and group-by queries on those dimensions can use skip indexes and avoid full segment scans. Materialized views pre-aggregate common query patterns at write time to serve dashboards with sub-second latency.

How does user identity stitching work in an event analytics pipeline?

User identity stitching is the process of associating events from anonymous sessions with a known user account after login or sign-up. Before login, events are tagged with an anonymous device or session ID. When the user authenticates, the client SDK sends an identify call that maps the anonymous ID to the authenticated user ID. The analytics backend creates an alias record linking both IDs, then retroactively re-keys historical anonymous events to the user ID, or stores the alias and resolves it at query time via a join. On the server side, stitching is typically handled by a customer data platform (CDP) or an identity resolution service that maintains a graph of all known IDs (email, device ID, cookie, user ID) for each individual, allowing accurate cross-device and cross-session attribution.

Low Level Design: Event Analytics Pipeline

⏱ 10 min read

Overview

An event analytics pipeline is the backbone of any product analytics product. It ingests raw behavioral events from client SDKs, buffers them reliably, processes and enriches them in a stream, stores them in a query-optimized OLAP store, and exposes a query API for dashboards and ad-hoc analysis. This low level design walks through every component with concrete data structures, API contracts, and scaling decisions.

Requirements

Functional

Accept events from web, iOS, and Android SDKs over HTTP.
Validate, enrich (geo-IP, user-agent parsing), and deduplicate events.
Support arbitrary event schemas defined by tenants (schema-on-read with optional schema enforcement).
Store events in an OLAP engine queryable by SQL.
Expose a query API that translates structured queries into SQL and returns results with pagination.
Retain raw events for 12 months; aggregated data indefinitely.

Non-functional

Ingest: 1 M events/sec sustained, 3 M/sec burst.
End-to-end latency (event visible in query results): < 30 s p99.
Query API p99 for last-7-day queries on a single tenant: < 2 s.
Durability: zero data loss after acknowledgment.

Event Schema Design

Canonical Event Envelope

All events, regardless of tenant-defined properties, share a canonical envelope:

{
  // System fields (set by collection layer, immutable)
  event_id:      "018f-...",          // UUIDv7 (time-ordered)
  tenant_id:     "acme",
  received_at:   "2024-04-13T10:00:00.123Z",
  sdk_version:   "js-2.4.1",
  ip:            "203.0.113.5",       // dropped after enrichment
  // Client fields (set by SDK)
  event_type:    "button_click",
  sent_at:       "2024-04-13T09:59:59.900Z",
  user_id:       "u_abc",
  anonymous_id:  "anon_xyz",
  session_id:    "sess_qrs",
  // Enriched fields (set by enrichment layer)
  country:       "US",
  region:        "CA",
  city:          "San Francisco",
  device_type:   "desktop",
  browser:       "Chrome",
  os:            "macOS",
  // Tenant-defined properties (arbitrary key-value)
  properties: {
    button_label: "Upgrade",
    page:         "/pricing",
    plan:         "pro"
  }
}

UUIDv7 for event_id

UUIDv7 encodes a millisecond timestamp in the high bits, making IDs time-ordered. This gives two benefits: deduplication state can be purged by time without a full scan, and OLAP engines can use the ID as a clustering key for efficient range scans.

Collection SDK

JavaScript SDK

The SDK is loaded as a snippet that bootstraps an async queue. Key behaviors:

Buffering: events are buffered in memory and flushed every 2 s or at 25 events.
Persistence: unflushed events are serialized to localStorage on visibilitychange (tab hide) so they survive page navigations.
Transport: primary transport is fetch with keepalive. Fallback is navigator.sendBeacon for unload events. Both use gzip-compressed JSON batches.
Retry: exponential backoff (1 s, 2 s, 4 s) on 5xx or network error. 429 responses respect the Retry-After header.
Identity: anonymous_id is a UUIDv4 persisted in a first-party cookie (SameSite=Lax, 2-year expiry). user_id is set via an explicit identify() call and joined server-side.

Mobile SDKs (iOS / Android)

Mobile SDKs use SQLite as a durable event queue. Events are written synchronously to SQLite before the enqueue call returns, guaranteeing no data loss even on app crash. A background worker dequeues and POSTs batches, deleting rows only after HTTP 200/202. Batches are capped at 500 KB to avoid request timeouts on slow mobile networks.

Ingestion Buffer

Collection Endpoint

Stateless Go services accept POST /v1/batch with a JSON array of events. Per-request processing:

Authenticate: validate API key against an in-memory LRU cache (1 000 tenants, 60 s TTL).
Parse: unmarshal JSON, reject malformed payloads with HTTP 400.
Stamp: set received_at = now(), derive server_id for tracing.
Size check: reject batches > 1 MB (HTTP 413).
Produce to Kafka topic events.raw, partition key = tenant_id.
Return HTTP 202 (async; no processing guarantee beyond Kafka durability).

The Kafka producer is configured with acks=all, enable.idempotence=true, compression.type=snappy, linger.ms=10, batch.size=256KB. This achieves > 98% compression on typical JSON event payloads and batches network I/O.

Kafka Topic Configuration

Topic: events.raw
  Partitions:         512
  Replication factor: 3
  Min ISR:            2
  Retention:          48 h (safety buffer; Flink processes within seconds)
  Segment size:       256 MB
  Partition key:      tenant_id (ensures per-tenant ordering)

Stream Processing

Enrichment Job (Flink)

The enrichment Flink job reads from events.raw and writes to events.enriched. Per-event operations (all asynchronous to avoid blocking the main pipeline thread):

Geo-IP lookup: async call to an in-process MaxMind GeoLite2-City database (memory-mapped, ~60 MB). Resolves country, region, city from IP. IP is then dropped.
User-agent parsing: ua-parser2 in-process library. Resolves device_type, browser, browser_version, os, os_version.
Identity resolution: async lookup of user_id → canonical_user_id mapping from a Redis hash (populated by identify() calls). Merges anonymous_id to canonical_user_id if a mapping exists.
Schema validation: optional per-tenant JSON Schema validation. Violations are routed to a dead-letter topic events.dlq rather than dropped.

Deduplication Job (Flink)

A separate Flink job reads from events.enriched and performs deduplication before writing to events.deduped.

State: RocksDB-backed keyed state on event_id, TTL = 1 hour (covers SDK retry windows). On each event:

If event_id exists in state: drop the event, increment a dedup_count metric.
If not: add to state, forward the event.

The 1-hour TTL means events arriving more than 1 hour late could produce duplicates. The OLAP storage layer handles this with an idempotent upsert keyed on event_id.

OLAP Storage

ClickHouse Schema

Raw events are stored in a ClickHouse distributed table:

CREATE TABLE events (
  tenant_id      LowCardinality(String),
  event_id       String,
  event_type     LowCardinality(String),
  received_at    DateTime64(3),
  sent_at        DateTime64(3),
  user_id        String,
  anonymous_id   String,
  session_id     String,
  country        LowCardinality(String),
  region         LowCardinality(String),
  device_type    LowCardinality(String),
  browser        LowCardinality(String),
  os             LowCardinality(String),
  properties     String              -- JSON blob
) ENGINE = ReplicatedReplacingMergeTree(received_at)
  PARTITION BY (tenant_id, toYYYYMM(received_at))
  ORDER BY (tenant_id, event_type, toStartOfHour(received_at), event_id)
  SETTINGS index_granularity = 8192;

ReplacingMergeTree (replacing version = received_at) handles late duplicate inserts: ClickHouse background merges keep only the row with the highest received_at per sort key, which for duplicate event_ids means the later (re-inserted) copy. Combined with ORDER BY including event_id, this achieves eventual deduplication at the storage layer.

The properties JSON blob is queried using ClickHouse's JSON path functions (JSONExtractString, JSONExtractFloat). For high-cardinality property columns accessed frequently, tenants can define materialized columns:

ALTER TABLE events ADD COLUMN plan LowCardinality(String)
  MATERIALIZED JSONExtractString(properties, 'plan');

ClickHouse Sink (Flink)

The Flink ClickHouse sink uses the official JDBC connector with batching: flush every 5 s or 10 000 rows, whichever comes first. Inserts use HTTP interface for better batching semantics. Checkpointing ensures exactly-once delivery at the Flink level; ClickHouse ReplacingMergeTree handles the storage-level idempotency.

Query API

Query Model

The query API accepts structured JSON queries, avoiding the need to expose raw SQL to tenants:

POST /v1/query
{
  "tenant_id":  "acme",
  "event_type": "button_click",
  "time_range": {
    "start": "2024-04-06T00:00:00Z",
    "end":   "2024-04-13T00:00:00Z"
  },
  "filters": [
    {field: "country",             op: "eq",       value: "US"},
    {field: "properties.plan",     op: "in",       value: ["pro","enterprise"]}
  ],
  "aggregations": [
    {function: "count",            alias: "total"},
    {function: "count_distinct",   field: "user_id", alias: "unique_users"}
  ],
  "group_by":   ["country", "device_type"],
  "order_by":   [{field: "total", direction: "desc"}],
  "limit":      100,
  "cursor":     null
}

SQL Translation

The query service translates the above into parameterized ClickHouse SQL:

SELECT
  country,
  device_type,
  count()                    AS total,
  uniq(user_id)              AS unique_users
FROM events
WHERE tenant_id = 'acme'
  AND event_type = 'button_click'
  AND received_at >= '2024-04-06 00:00:00'
  AND received_at <  '2024-04-13 00:00:00'
  AND country = 'US'
  AND JSONExtractString(properties, 'plan') IN ('pro','enterprise')
GROUP BY country, device_type
ORDER BY total DESC
LIMIT 101;  -- fetch 101 to detect if there is a next page

Cursor-based pagination encodes the last row's sort key values. The next page query adds a WHERE clause on those values, avoiding the OFFSET performance problem in ClickHouse.

Query Timeout and Circuit Breaker

All ClickHouse queries run with max_execution_time=10 (seconds). The query service imposes a 15 s application-level timeout. A per-tenant circuit breaker (Hystrix-style) opens after 5 consecutive timeouts, returning HTTP 503 with a Retry-After header, preventing a runaway query from consuming all ClickHouse threads.

Identity Resolution

When a user identifies (logs in), the SDK calls analytics.identify(userId). This produces an identify event stored in a separate Redis hash: HSET identity:{tenant} {anonymous_id} {user_id}. The enrichment Flink job looks up this mapping for each event, backfilling user_id on events that previously had only anonymous_id. Historical event backfill (retroactively assigning user_id to past anonymous events) is handled by a nightly batch Spark job that rewrites affected ClickHouse partitions.

Data Retention and TTL

-- Raw events: 12-month TTL
ALTER TABLE events MODIFY TTL received_at + INTERVAL 12 MONTH DELETE;
-- Rollup tables: no TTL (retain indefinitely)
-- Kafka events.raw: 48 h
-- Kafka events.enriched, events.deduped: 6 h

Monitoring and Observability

Collection lag: Kafka consumer group lag for Flink enrichment and dedup jobs. Alert if lag > 100 K events.
End-to-end latency: a synthetic event injected every 30 s with a known event_id. Query API polls until it appears. Alert if latency > 30 s.
DLQ depth: events.dlq message count. Alert on any new messages (indicates schema violations or enrichment failures).
ClickHouse merge depth: number of parts per partition. Alert if parts > 300 (indicates insert rate exceeding merge rate).

Interview Tips

Be ready to discuss why deduplication happens in Flink (low latency, exact dedup within retry window) AND in ClickHouse (eventual correctness for late arrivals).
Explain the trade-off between schema-on-read (flexible, slower queries on properties JSON) vs materialized columns (fast, requires schema decisions upfront).
Know that ClickHouse ReplacingMergeTree deduplication is eventual — reads may see duplicates until a merge happens. Use FINAL modifier or deduplicate at query time for correctness-critical queries.
Discuss identity stitching complexity: the retroactive backfill problem and why it requires a batch rewrite rather than a streaming update.