Twitter (X) Data Engineer Interview Guide

X's data engineering org runs lean. The day-in-life data for this role shows a single engineer bouncing between Dataflow OOM debugging, PySpark pipeline builds, Airflow DAG refactors, BigQuery optimization, and on-call handoffs, all in one week. If you're coming from a company where pipeline work is split across ingestion, transformation, and platform teams, expect a much wider ownership surface here.

Twitter (X) Data Engineer Role

Success after year one means you reliably own a pipeline domain end-to-end, from Kafka event ingestion through processing (Dataflow, Spark) to BigQuery serving layers, with SLAs that hold without someone else backstopping you. Based on the day-in-life data, you'll be building ad impression attribution jobs, tuning feature store refresh queries for the For You timeline ranking, and writing migration plans for Trust & Safety data consumers. New ingestion paths appear as the product surface expands, and you're expected to design those paths, not just implement a spec.

A Typical Week

The split between infrastructure work and pure coding is closer than you'd expect at most companies. What that looks like in practice: Monday morning you're bumping Dataflow worker memory after a quote-tweet volume spike caused an OOM, and by Thursday afternoon you're writing pytest fixtures for the PySpark attribution job you've been building all week. The writing allocation is worth paying attention to, because design docs, migration plans, and runbook updates aren't optional when you may be the only person who knows why a particular DAG exists.

Projects & Impact Areas

Ad impression attribution stands out as the highest-visibility pipeline work in the day-in-life data. It joins Kafka ad impression streams with conversion events in GCS using a 7-day lookback window, and latency regressions here affect bid optimization directly. Alongside that revenue-critical work, you'll find yourself cleaning up 14TB of orphaned GCS Parquet files from deprecated pipelines, prototyping Apache Iceberg migrations for high-volume event tables, and writing schema migration plans so the Trust & Safety team can get denormalized action-level data instead of session-grain rollups.

Skills & What's Expected

Production-grade coding is the most underrated skill for this role. Candidates over-index on knowing Spark APIs and under-index on writing Python or Scala that handles edge cases, includes tests, and survives code review from engineers who ship services. The medium rating on ML/GenAI might tempt you to skip it, but the day-in-life shows you'll be tuning BigQuery queries that feed the ML feature store for timeline ranking, so you need to reason about how ML teams consume your output. Cloud infrastructure knowledge matters more here than at many data engineering roles because the stack is actively GCP-native (Dataflow, BigQuery, GCS), and you're expected to make cost and performance decisions like switching from on-demand to reserved slot pools.

Levels & Career Growth

The widget shows the level bands, but here's what the data implies about the gaps between them. At L3-L4 you own individual pipelines and prove reliability. At L5 you own a domain (say, all ad-event ingestion) and set SLAs. The jump to L6 is where people stall, because it's not about writing better Spark jobs. It's about defining cross-team data platform patterns that other engineers adopt without you pushing them. Demonstrating measurable reliability improvements (SLO attainment, incident reduction) is the strongest signal for advancing above L5.

Work Culture

The culture notes indicate X expects in-office presence at San Francisco HQ most days, though some teams have reportedly negotiated hybrid arrangements depending on manager discretion. On-call weeks can be intense given the lean headcount absorbing the same event volume, and the day-in-life data reflects that reality: weekend SLA breaches, OOM failures to triage Monday morning, and runbook handoffs every Friday. On the positive side, the team structure means less bureaucracy between you and the people who use your pipelines, and your work is visible to leadership without layers of middle management filtering it.

Twitter (X) Data Engineer Compensation

X is a private company, so any equity in your offer is illiquid. From what candidates report, that changes the math dramatically at senior levels, where the stock component can represent a third or more of headline total comp. Until there's a liquidity event, you can't spend those shares. Get every detail of your grant in writing before signing: vesting cadence, cliff, refresh policy, and any performance conditions. The source data on X's specific equity mechanics is thin, so treat anything a recruiter tells you verbally as unconfirmed until it's in your offer letter.

Your strongest negotiation levers, per the offer notes, are equity amount, sign-on bonus, and level. Level is worth special attention because moving up even one rung shifts every comp component (base, stock, bonus) into a higher band simultaneously. If you're carrying a competing offer from a public company where RSUs convert to cash quarterly, lean on that contrast to push for a larger sign-on bonus or higher base, since those are the parts of the package you can actually deposit.

Twitter (X) Data Engineer Interview Process

The loop runs about four weeks from recruiter call to offer. System design is the round where, from what candidates report, rejection rates are highest. X's interviewers expect you to reason about pipeline architectures for their specific product surfaces (ad auction events, timeline engagement streams, trust/safety classification feeds) and to address operational realities like schema evolution with Protobuf, replay strategies against Manhattan, and monitoring for Kafka consumer lag. Vague component diagrams won't cut it.

Don't treat the Hiring Manager Screen as a warmup. The round description makes clear the HM is probing whether you've owned pipelines end-to-end, from defining SLAs through incident response, not just whether you've touched Spark or Airflow. Come ready to walk through a specific architecture you built, the tradeoffs you chose (batch vs. streaming, cost vs. freshness), and the measurable outcomes you delivered.

Twitter (X) Data Engineer Interview Questions

Data Pipelines & Engineering

Expect questions that force you to design reliable batch/streaming flows for training and online features (e.g., Kafka/Flink + Airflow/Dagster). You’ll be evaluated on backfills, late data, idempotency, SLAs, lineage, and operational failure modes.

What is the difference between a batch pipeline and a streaming pipeline, and when would you choose each?

EasyFundamentals

Sample Answer

Batch pipelines process data in scheduled chunks (e.g., hourly, daily ETL jobs). Streaming pipelines process data continuously as it arrives (e.g., Kafka + Flink). Choose batch when: latency tolerance is hours or days (daily reports, model retraining), data volumes are large but infrequent, and simplicity matters. Choose streaming when you need real-time or near-real-time results (fraud detection, live dashboards, recommendation updates). Most companies use both: streaming for time-sensitive operations and batch for heavy analytical workloads, model training, and historical backfills.

You ingest Kafka events for booking state changes (created, confirmed, canceled) into a Hive table, then daily compute confirmed_nights per listing for search ranking. How do you make the Spark job idempotent under retries and late-arriving cancels without double counting?

AirbnbMediumIdempotency and Late Data

Sample Answer

Most candidates default to append-only aggregations with a checkpoint, but that fails here because duplicates and late cancels mutate history and you will overcount. You need a stable event key (booking_id plus event_version or event_timestamp plus source offset) and a dedupe rule, then compute state using last-write-wins or a state machine per booking. Write results with upserts (partition overwrite, MERGE, or Hudi/Iceberg style) keyed by booking_id and listing_id, and drive the aggregation off the canonical booking state, not raw events. Add watermarking and a correction window so late cancels trigger targeted recompute instead of full backfills.

You need a pipeline that produces a near real-time host payout ledger: streaming updates every minute, but also a daily audited snapshot that exactly matches finance when late adjustments arrive up to 30 days. Design the batch plus streaming architecture, including how you handle schema evolution and backfills without breaking downstream tables.

AirbnbHardLambda Architecture and Backfills

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Data Pipelines & Engineering questions

System Design

Most candidates underestimate how much your design must balance latency, consistency, and cost at top tech companies scale. You’ll be evaluated on clear component boundaries, failure modes, and how you’d monitor and evolve the system over time.

Design a dataset registry for LLM training and evaluation that lets you reproduce any run months later, including the exact prompt template, filtering rules, and source snapshots. What metadata and storage layout do you require, and which failure modes does it prevent?

AnthropicMediumDataset Versioning and Lineage

Sample Answer

Use an immutable, content-addressed dataset registry that writes every dataset as a manifest of exact source pointers, transforms, and hashes, plus a separate human-readable release record. Store raw sources append-only, store derived datasets as partitioned files keyed by dataset_id and version, and capture code commit SHA, config, and schema in the manifest so reruns cannot drift. This prevents silent data changes, schema drift, and accidental reuse of a similarly named dataset, which is where most people fail.

A company wants a unified fact table for Marketplace Orders (bookings, cancellations, refunds, chargebacks) that supports finance reporting and ML features, while source systems emit out-of-order updates and occasional duplicates. Design the data model and pipeline, including how you handle upserts, immutable history, backfills, and data quality gates at petabyte scale.

AirbnbHardWarehouse Modeling and Incremental Processing

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more System Design questions

SQL & Data Manipulation

Your SQL will get stress-tested on joins, window functions, deduping, and incremental logic that mirrors real ETL/ELT work. Common pitfalls include incorrect grain, accidental fan-outs, and filtering at the wrong stage.

Airflow runs a daily ETL that builds fact_host_daily(host_id, ds, active_listings, booked_nights). Source tables are listings(listing_id, host_id, created_at, deactivated_at) and bookings(booking_id, listing_id, check_in, check_out, status, created_at, updated_at). Write an incremental SQL for ds = :run_date that counts active_listings at end of day and booked_nights for stays overlapping ds, handling late-arriving booking updates by using updated_at.

AirbnbMediumIncremental ETL and Late Arriving Data

Sample Answer

Walk through the logic step by step as if thinking out loud. You start by defining the day window, ds start and ds end. Next, active_listings is a snapshot metric, so you count listings where created_at is before ds end, and deactivated_at is null or after ds end. Then booked_nights is an overlap metric, so you compute the intersection of [check_in, check_out) with [ds, ds+1), but only for non-canceled bookings. Finally, for incrementality you only scan bookings that could affect ds, either the stay overlaps ds or the record was updated recently, and you upsert the single ds partition for each host.

SQLCopy

1WITH params AS (
2  SELECT
3    CAST(:run_date AS DATE) AS ds,
4    CAST(:run_date AS TIMESTAMP) AS ds_start_ts,
5    CAST(:run_date AS TIMESTAMP) + INTERVAL '1' DAY AS ds_end_ts
6),
7active_listings_by_host AS (
8  SELECT
9    l.host_id,
10    p.ds,
11    COUNT(*) AS active_listings
12  FROM listings l
13  CROSS JOIN params p
14  WHERE l.created_at < p.ds_end_ts
15    AND (l.deactivated_at IS NULL OR l.deactivated_at >= p.ds_end_ts)
16  GROUP BY l.host_id, p.ds
17),
18-- Limit booking scan for incremental run.
19-- Assumption: you run daily and keep a small lookback for late updates.
20-- This reduces IO while still catching updates that change ds attribution.
21bookings_candidates AS (
22  SELECT
23    b.booking_id,
24    b.listing_id,
25    b.check_in,
26    b.check_out,
27    b.status,
28    b.updated_at
29  FROM bookings b
30  CROSS JOIN params p
31  WHERE b.updated_at >= p.ds_start_ts - INTERVAL '7' DAY
32    AND b.updated_at < p.ds_end_ts + INTERVAL '1' DAY
33),
34booked_nights_by_host AS (
35  SELECT
36    l.host_id,
37    p.ds,
38    SUM(
39      CASE
40        WHEN bc.status = 'canceled' THEN 0
41        -- Compute overlap nights between [check_in, check_out) and [ds, ds+1)
42        ELSE GREATEST(
43          0,
44          DATE_DIFF(
45            'day',
46            GREATEST(CAST(bc.check_in AS DATE), p.ds),
47            LEAST(CAST(bc.check_out AS DATE), p.ds + INTERVAL '1' DAY)
48          )
49        )
50      END
51    ) AS booked_nights
52  FROM bookings_candidates bc
53  JOIN listings l
54    ON l.listing_id = bc.listing_id
55  CROSS JOIN params p
56  WHERE CAST(bc.check_in AS DATE) < p.ds + INTERVAL '1' DAY
57    AND CAST(bc.check_out AS DATE) > p.ds
58  GROUP BY l.host_id, p.ds
59),
60final AS (
61  SELECT
62    COALESCE(al.host_id, bn.host_id) AS host_id,
63    (SELECT ds FROM params) AS ds,
64    COALESCE(al.active_listings, 0) AS active_listings,
65    COALESCE(bn.booked_nights, 0) AS booked_nights
66  FROM active_listings_by_host al
67  FULL OUTER JOIN booked_nights_by_host bn
68    ON bn.host_id = al.host_id
69   AND bn.ds = al.ds
70)
71-- In production this would be an upsert into the ds partition.
72SELECT *
73FROM final
74ORDER BY host_id;

Event stream table listing_price_events(listing_id, event_time, ingest_time, price_usd) can contain duplicates and out-of-order arrivals. Write SQL to build a daily snapshot table listing_price_daily(listing_id, ds, price_usd, event_time) for ds = :run_date using the latest event_time within the day, breaking ties by latest ingest_time, and ensuring exactly one row per listing per ds.

AirbnbHardWindow Functions and Deduping

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more SQL & Data Manipulation questions

Data Warehouse

A the company client wants one the company account shared by 15 business units, each with its own analysts, plus a central the company X delivery team that runs dbt and Airflow. Design the warehouse layer and access model (schemas, roles, row level security, data products) so units cannot see each other’s data but can consume shared conformed dimensions.

Boston Consulting Group (BCG)MediumMulti-tenant warehouse architecture and access control

Sample Answer

Most candidates default to separate databases per business unit, but that fails here because conformed dimensions and shared transformation code become duplicated and drift fast. You want a shared curated layer for conformed entities (customer, product, calendar) owned by a platform team, plus per unit marts or data products with strict role based access control. Use the company roles with least privilege, database roles, and row access policies (and masking policies) keyed on tenant identifiers where physical separation is not feasible. Put ownership, SLAs, and contract tests on the shared layer so every unit trusts the same definitions.

A Redshift cluster powers an operations dashboard where 150 concurrent users run the same 3 queries, one query scans fact_clickstream (10 TB) joined to dim_sku and dim_marketplace and groups by day and marketplace, but it spikes to 40 minutes at peak. What concrete Redshift table design changes (DISTKEY, SORTKEY, compression, materialized views) and workload controls would you apply, and how do you validate each change with evidence?

AmazonHardRedshift Physical Design and Concurrency

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Data Warehouse questions

Data Modeling

Rather than raw SQL skill, you’re judged on how you structure facts, dimensions, and metrics so downstream analytics stays stable. Watch for prompts around SCD types, grain definition, and metric consistency across Sales/Analytics consumers.

A company has a daily snapshot table listing_snapshot(listing_id, ds, price, is_available, host_id, city_id) and an events table booking_event(booking_id, listing_id, created_at, check_in, check_out). Write SQL to compute booked nights and average snapshot price at booking time by city and ds, where snapshot ds is the booking created_at date.

AirbnbMediumSnapshot vs Event Join

Sample Answer

Start with what the interviewer is really testing: "This question is checking whether you can align event time to snapshot time without creating fanout joins or time leakage." You join booking_event to listing_snapshot on listing_id plus the derived snapshot date, then aggregate nights as $\text{datediff}(\text{check\_out}, \text{check\_in})$. You also group by snapshot ds and city_id, and you keep the join predicates tight so each booking hits at most one snapshot row.

SQLCopy

1SELECT
2  ls.ds,
3  ls.city_id,
4  SUM(DATE_DIFF('day', be.check_in, be.check_out)) AS booked_nights,
5  AVG(ls.price) AS avg_snapshot_price_at_booking
6FROM booking_event be
7JOIN listing_snapshot ls
8  ON ls.listing_id = be.listing_id
9 AND ls.ds = DATE(be.created_at)
10GROUP BY 1, 2;

You are designing a star schema for host earnings and need to support two use cases: monthly payouts reporting and real-time fraud monitoring on payout anomalies. How do you model payout facts and host and listing dimensions, including slowly changing attributes like host country and payout method, so both use cases stay correct?

AirbnbHardSCD and Multi-Use Fact Modeling

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Data Modeling questions

Coding & Algorithms

Your ability to reason about constraints and produce correct, readable Python under time pressure is a major differentiator. You’ll need solid data-structure choices, edge-case handling, and complexity awareness rather than exotic CS theory.

Given a stream of (asin, customer_id, ts) clicks for an detail page, compute the top K ASINs by unique customer count within the last 24 hours for a given query time ts_now. Input can be unsorted, and you must handle duplicates and out-of-window events correctly.

AmazonMediumSliding Window Top-K

Sample Answer

Get this wrong in production and your top ASIN dashboard flaps, because late events and duplicates inflate counts and reorder the top K every refresh. The right call is to filter by the $24$ hour window relative to ts_now, dedupe by (asin, customer_id), then use a heap or partial sort to extract K efficiently.

PythonCopy

1from __future__ import annotations
2
3from datetime import datetime, timedelta
4from typing import Iterable, List, Tuple, Dict, Set
5import heapq
6
7
8def _parse_time(ts: str) -> datetime:
9    """Parse ISO-8601 timestamps, supporting a trailing 'Z'."""
10    if ts.endswith("Z"):
11        ts = ts[:-1] + "+00:00"
12    return datetime.fromisoformat(ts)
13
14
15def top_k_asins_unique_customers_last_24h(
16    events: Iterable[Tuple[str, str, str]],
17    ts_now: str,
18    k: int,
19) -> List[Tuple[str, int]]:
20    """Return top K (asin, unique_customer_count) in the last 24h window.
21
22    events: iterable of (asin, customer_id, ts) where ts is ISO-8601 string.
23    ts_now: window reference time (ISO-8601).
24    k: number of ASINs to return.
25
26    Ties are broken by ASIN lexicographic order (stable, deterministic output).
27    """
28    now = _parse_time(ts_now)
29    start = now - timedelta(hours=24)
30
31    # Deduplicate by (asin, customer_id) within the window.
32    # If events are huge, you would partition by asin or approximate, but here keep it exact.
33    seen_pairs: Set[Tuple[str, str]] = set()
34    customers_by_asin: Dict[str, Set[str]] = {}
35
36    for asin, customer_id, ts in events:
37        t = _parse_time(ts)
38        if t < start or t > now:
39            continue
40        pair = (asin, customer_id)
41        if pair in seen_pairs:
42            continue
43        seen_pairs.add(pair)
44        customers_by_asin.setdefault(asin, set()).add(customer_id)
45
46    # Build counts.
47    counts: List[Tuple[int, str]] = []
48    for asin, custs in customers_by_asin.items():
49        counts.append((len(custs), asin))
50
51    if k <= 0:
52        return []
53
54    # Get top K by count desc, then asin asc.
55    # heapq.nlargest uses the tuple ordering, so use (count, -) carefully.
56    top = heapq.nlargest(k, counts, key=lambda x: (x[0], -ord(x[1][0]) if x[1] else 0))
57
58    # The key above is not a correct general lexicographic tiebreak, so do it explicitly.
59    # Sort all candidates by (-count, asin) and slice K. This is acceptable for moderate cardinality.
60    top_sorted = sorted(((asin, cnt) for cnt, asin in counts), key=lambda p: (-p[1], p[0]))
61    return top_sorted[:k]
62
63
64if __name__ == "__main__":
65    data = [
66        ("B001", "C1", "2024-01-02T00:00:00Z"),
67        ("B001", "C1", "2024-01-02T00:01:00Z"),  # duplicate customer for same ASIN
68        ("B001", "C2", "2024-01-02T01:00:00Z"),
69        ("B002", "C3", "2024-01-01T02:00:00Z"),
70        ("B003", "C4", "2023-12-31T00:00:00Z"),  # out of window
71    ]
72    print(top_k_asins_unique_customers_last_24h(data, "2024-01-02T02:00:00Z", 2))
73

Given a list of nightly booking records {"listing_id": int, "guest_id": int, "checkin": int day, "checkout": int day} (checkout is exclusive), flag each listing_id that is overbooked, meaning at least one day has more than $k$ active stays, and return the earliest day where the maximum occupancy exceeds $k$.

AirbnbHardSweep Line Overlap Counting

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Coding & Algorithms questions

Data Engineering

You need to join a 5 TB Delta table of per-frame telemetry with a 50 GB Delta table of trip metadata on trip_id to produce a canonical fact table in the company. Would you rely on broadcast join or shuffle join, and what explicit configs or hints would you set to make it stable and cost efficient?

CruiseMediumSpark Joins and Partitioning

Sample Answer

You could force a broadcast join of the 50 GB table or run a standard shuffle join on trip_id. Broadcast wins only if the metadata table can reliably fit in executor memory across the cluster, otherwise you get OOM or repeated GC and retries. In most real clusters 50 GB is too big to broadcast safely, so shuffle join wins, then you make it stable by pre-partitioning or bucketing by trip_id where feasible, tuning shuffle partitions, and enabling AQE to coalesce partitions.

PythonCopy

1from pyspark.sql import functions as F
2
3# Inputs
4telemetry = spark.read.format("delta").table("raw.telemetry_frames")  # very large
5trips = spark.read.format("delta").table("dim.trip_metadata")          # large but smaller
6
7# Prefer shuffle join with AQE for stability
8spark.conf.set("spark.sql.adaptive.enabled", "true")
9spark.conf.set("spark.sql.adaptive.coalescePartitions.enabled", "true")
10
11# Right-size shuffle partitions, set via env or job config in practice
12spark.conf.set("spark.sql.shuffle.partitions", "4000")
13
14# Pre-filter early if possible to reduce shuffle
15telemetry_f = telemetry.where(F.col("event_date") >= F.date_sub(F.current_date(), 7))
16trips_f = trips.select("trip_id", "vehicle_id", "route_id", "start_ts", "end_ts")
17
18joined = (
19    telemetry_f
20    .join(trips_f.hint("shuffle_hash"), on="trip_id", how="inner")
21)
22
23# Write out with sane partitioning and file sizing
24(
25    joined
26    .repartition("event_date")
27    .write
28    .format("delta")
29    .mode("overwrite")
30    .option("overwriteSchema", "true")
31    .saveAsTable("canon.fact_telemetry_enriched")
32)

A company Support wants a governed semantic layer for "First Response Time" and "Resolution Time" across email and chat, and an LLM tool will answer questions using those metrics. How do you enforce metric definitions, data access, and quality guarantees so the LLM and Looker both return consistent numbers and do not leak restricted fields?

CoinbaseHardGovernance, Semantic Layer, and Access Controls

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Data Engineering questions

Cloud Infrastructure

In practice, you’ll need to articulate why you’d pick Spark/Hive vs an MPP warehouse vs Cassandra for a specific workload. Interviewers look for pragmatic tradeoffs: throughput vs latency, partitioning/sharding choices, and operational constraints.

A the company warehouse for a client’s KPI dashboard has unpredictable concurrency, and monthly spend is spiking. What specific changes do you make to balance performance and cost, and what signals do you monitor to validate the change?

Boston Consulting Group (BCG)MediumCost and performance optimization

Sample Answer

The standard move is to right-size compute, enable auto-suspend and auto-resume, and separate workloads with different warehouses (ELT, BI, ad hoc). But here, concurrency matters because scaling up can be cheaper than scaling out if query runtime drops sharply, and scaling out can be required if queueing dominates. You should call out monitoring of queued time, warehouse load, query history, cache hit rates, and top cost drivers by user, role, and query pattern. You should also mention guardrails like resource monitors and workload isolation via roles and warehouse assignment.

You need near real-time order events (p95 under 5 seconds) for an Operations dashboard and also a durable replayable history for backfills, events are 20k per second at peak. How do you choose between Kinesis Data Streams plus Lambda versus Kinesis Firehose into S3 plus Glue, and what IAM, encryption, and monitoring controls do you put in place?

AmazonHardStreaming Ingestion Tradeoffs and Security

Sample Answer

Unlock the full solution

View detailed answers and practice in our question bank

View Solution

Practice more Cloud Infrastructure questions

X's loop is weighted toward the unsexy, high-stakes work that actually keeps the platform running: designing Kafka-to-warehouse ingestion for hundreds of billions of daily events, then proving you can query and model that data under real constraints like partition skew on tweet_id or SCD handling when post schemas evolve with new fields like community_id. The compounding trap is that pipeline design and SQL questions both assume you already think in terms of idempotent backfills, partitioning strategy, and SLA-aware orchestration, so a gap in either area undermines your answers in the other. Meanwhile, the coding slice focuses on data engineering glue (schema validation, deduplication, streaming-safe counters), not algorithmic puzzles, which means drilling competitive programming problems is probably the lowest-ROI way to spend your prep time.

Practice with pipeline, SQL, and modeling questions tailored to roles like this at datainterview.com/questions.

How to Prepare for Twitter (X) Data Engineer Interviews

X's stated ambition is to become an "everything app" that integrates payments, banking, video, and messaging into a single platform. For data engineers, that means the job isn't maintaining a mature social media pipeline. It's building net-new data infrastructure for verticals that didn't exist at Twitter two years ago, while keeping ad revenue pipelines healthy given $2.5B in annual revenue declining at -13.8% year-over-year.

The "why X?" answer that falls flat is some variation of "I believe in the everything-app vision." What separates strong candidates is demonstrating you've studied X's actual tech stack and can speak to specific engineering tradeoffs. Reference the Kafka-based event infrastructure, the reported cloud migration challenges, or how X's history of building custom internal tools rather than buying off the shelf shapes what a DE owns day to day. That kind of answer shows you've done homework beyond the headlines.

Try a Real Interview Question

1WITH latest_event AS (
2  SELECT
3    e.transaction_id,
4    e.merchant_id,
5    e.event_ts,
6    DATE(e.event_ts) AS event_date,
7    e.status,
8    CAST(e.amount_usd AS NUMERIC) AS amount_usd,
9    ROW_NUMBER() OVER (
10      PARTITION BY e.transaction_id
11      ORDER BY e.event_ts DESC
12    ) AS rn
13  FROM payment_events e
14  WHERE DATE(e.event_ts) BETWEEN DATE '2026-01-10' AND DATE '2026-01-11'
15)
16SELECT
17  le.event_date,
18  le.merchant_id,
19  SUM(
20    CASE
21      WHEN le.status = 'COMPLETED' THEN le.amount_usd
22      WHEN le.status IN ('REFUNDED', 'CHARGEBACK') THEN -le.amount_usd
23      ELSE CAST(0 AS NUMERIC)
24    END
25  ) AS net_amount_usd
26FROM latest_event le
27WHERE le.rn = 1
28GROUP BY le.event_date, le.merchant_id
29ORDER BY le.event_date, le.merchant_id;

X's coding rounds reportedly emphasize production-quality data manipulation over algorithmic puzzles, which tracks with a post-layoff org where DEs ship code directly to pipelines serving ad auctions and content ranking. The problem above is a good proxy for that bar. Build your muscle memory with more problems like it at datainterview.com/coding.

Test Your Readiness

Spot your weak areas across X's question distribution, then drill them at datainterview.com/questions.