Airbyte vs Fivetran on a Data Engineering interview

Why this question shows up

When a Data Engineering interviewer at Snowflake, Databricks, Stripe or DoorDash asks "how do you load data from Postgres into the warehouse?", they are not testing whether you have memorized 400 connector names. They want to see if you understand the build-vs-buy decision, the per-row pricing trap, and how schema evolution breaks pipelines at 2am. The EL layer is the cheapest part of your stack until it isn't — and the moment a 50M-row table starts replicating hourly, the bill or the on-call burden explodes.

The honest version of this question is "have you ever been responsible for a managed-EL invoice, or for restarting an Airbyte pod at midnight?" Most candidates default to a marketing-page comparison: Airbyte is open source, Fivetran is managed, the end. That answer gets you a polite nod and a follow-up that exposes the gap. The strong answer talks about MAR-based pricing, CDC slot lag, connector quality variance, and the fact that the right tool depends on how many engineers you have, not how many connectors the vendor lists.

Airbyte in one breath

Airbyte is an open-source EL platform with a self-hosted option and a managed cloud. It ships 350+ connectors, most of them community-maintained, and lets you build custom ones using its connector development kit (CDK in Python or low-code YAML). The core sync modes are full refresh, incremental cursor-based, and CDC for popular databases (Postgres via logical replication, MySQL via binlog, MongoDB via oplog).

The pricing story has two paths. Self-hosted is free as in beer — you pay for the Kubernetes cluster, the storage, and the on-call rotation. Airbyte Cloud charges by MAR (monthly active rows), roughly $10 per million rows for low volumes and discounted at scale. The catch is operational maturity: connector quality varies wildly between the top 30 (Stripe, Salesforce, Postgres — well-maintained) and the long tail (random SaaS APIs maintained by one contributor in 2023).

Pick Airbyte when you have a platform team that can absorb the operational cost, or when your data has long-tail sources that managed vendors don't cover.

Fivetran in one breath

Fivetran is the managed-SaaS premium tier of the EL market. It offers 400+ enterprise-grade connectors, automatic schema evolution, 24/7 support, SOC 2 / HIPAA compliance, and a strong "just works" reputation. You point it at a source, point it at a destination, and it handles the rest — including connector breakages, schema changes, and replication-slot management.

Pricing is per MAR with credit-based tiers. For a typical mid-size company replicating Salesforce, Stripe, and a Postgres OLTP database, you should expect $3k-$15k per month. For high-volume sources — clickstream tables, large transactional databases — the bill can climb to $50k-$200k/month before you notice. The classic horror story is a events table set to incremental but not deduplicated, where a single backfill triggers a five-figure invoice.

Stitch, Meltano, dlt and friends

The middle of the market is more crowded than it looks. Stitch (Talend) is the budget alternative to Fivetran — fewer connectors (around 130), lower price, slimmer support. It's a reasonable choice for a startup with a Postgres-to-Snowflake pipeline and a tight budget.

Meltano is an open-source orchestrator built around the Singer protocol, the original tap-target spec from Stitch. It's code-first and Git-friendly, but the Singer connector ecosystem is uneven — many taps are abandoned. dlt (data load tool) is the new entrant: Python-native, code-first, declarative. Engineers love it because the pipeline is just a Python function, not a YAML mystery box. Hevo Data and Portable are second-tier managed options with niche connector coverage.

Then there's custom Python. Many serious teams write their own loaders with psycopg2, boto3, and Airflow. This sounds barbaric, but for a single high-volume source it's often the cheapest and most reliable option — you control every retry, every batch size, every dedup key.

# Minimal custom EL pattern
import psycopg2, boto3, json
from datetime import datetime, timedelta

def extract_orders(since: datetime):
    conn = psycopg2.connect(dsn=PG_DSN)
    cur = conn.cursor(name="server_cursor")
    cur.itersize = 10_000
    cur.execute(
        "SELECT * FROM orders WHERE updated_at > %s",
        (since,),
    )
    for row in cur:
        yield row

def load_to_s3(rows, key):
    s3 = boto3.client("s3")
    body = "\n".join(json.dumps(r, default=str) for r in rows)
    s3.put_object(Bucket="raw", Key=key, Body=body)

The point of the snippet is not that you should write this yourself — it's that a Data Engineer should be able to estimate when this 30-line script becomes cheaper than a managed connector.

Side-by-side comparison

Read this table as a decision aid, not a leaderboard — every row trades on a different axis.

How to answer the trade-off question

The interviewer asks: "You're hired as the first Data Engineer at a Series A. Pick an EL stack." The wrong answer names a tool first. The right answer asks two questions and then names a tool.

Question one: how many data sources, and how custom are they? If it's three to five mainstream SaaS sources (Stripe, Salesforce, HubSpot, Postgres) and one team to maintain everything, Fivetran is the right answer despite the cost — one tired engineer is more expensive than $5k/month of connectors. If it's twelve obscure sources and a platform team of three, self-hosted Airbyte or dlt wins.

Question two: what is the volume in MAR, and how fast is it growing? Per-row pricing is linear in MAR but the cost curve hits a knee around 50M MAR/month, where managed vendors stop being cheaper than a platform engineer. Forecast volume 18 months out, not today. A 5M-row table doubling every six months will be 80M in 18 months — and the Fivetran bill goes from $3k to $30k in the same window.

The answer that lands in an interview sounds like: "For three to five mainstream sources at sub-50M MAR with no dedicated platform team, I'd start on Fivetran or Airbyte Cloud — the operational cost dominates the license cost. If we cross 50M MAR or pick up custom sources, I'd migrate the heaviest tables to Airbyte self-hosted or custom Python, and keep the long tail on the managed tool. It's a portfolio, not a religion."

Common pitfalls

The first trap is self-hosting Airbyte on a single instance with no HA. The platform looks free until the EC2 box reboots during a CDC sync and the replication slot fills up the source database's WAL directory. The fix is unglamorous: deploy on Kubernetes with proper resource limits, run a separate worker pool from the scheduler, and monitor both pod restarts and source-side replication lag. Single-instance Airbyte is a demo, not a production deployment.

The second trap is adopting Fivetran without modeling the MAR bill on your worst table. Sales engineers will quote you a low base tier, but if you have a wide events table with frequent updates, the MAR count balloons. Before signing, run a one-week trial on your top three tables and extrapolate. The companies that get burned are the ones that priced the pipeline on the Salesforce volume and ignored the Postgres events table.

The third trap is writing custom Python pipelines for cases the connector market already solved. If your source is Stripe or Salesforce, the managed connector handles pagination, rate limits, webhook backfills, OAuth refresh, and schema drift — all things you will rediscover painfully on your own. Reach for custom code only when (a) the source is niche, or (b) the volume makes managed cost-prohibitive.

The fourth trap is not monitoring sync delays. A pipeline that silently falls 8 hours behind is worse than a pipeline that fails loudly. Set up freshness checks on the destination — MAX(updated_at) per table — and page someone when it crosses 2x the expected sync interval. The managed tools have built-in alerts but they are off by default for most metrics.

The fifth trap is the schema evolution surprise. The source team changes a column from INT to BIGINT, or renames user_id to customer_id, and your pipeline either fails or — worse — silently truncates. Airbyte and Fivetran both handle most additive changes, but destructive changes (type narrowing, column drops) need explicit handling. Set up alerts on connector configuration changes and treat schema drift as a code-review event, not a runtime event.

Related reading

• Airflow on a Data Engineering interview
• dbt incremental models on a DE interview
• Kafka Connect on a DE interview
• Merge and upsert patterns for DE
• SQL antipatterns on a DE interview

If you want to drill EL trade-off questions like this every day, NAILDD is launching with 500+ Data Engineering problems across exactly this pattern.

FAQ

Is Airbyte production-ready for large companies?

Yes for medium volumes, more carefully for petabyte-scale. Most users running Airbyte in production are in the 1M-100M MAR/month range across mixed sources. Above that, teams typically split: keep Airbyte for the long tail of SaaS sources and move the largest tables to custom CDC pipelines on Debezium + Kafka, or to managed CDC products like Striim or Estuary. The constraint isn't Airbyte itself — it's the connector quality variance, which hurts more when a single table is half your warehouse.

When is Fivetran clearly worth the price?

When the engineer-hours saved exceed the license cost. A senior DE fully loaded costs $200k-$300k/year. Fivetran at $3k-$10k/month replaces meaningful chunks of their on-call burden on standard SaaS sources. The math flips when you have a platform team already maintaining infrastructure, or when MAR is high enough that the bill is closer to a full FTE than to a side budget.

How does CDC fit into this picture?

CDC (change data capture) is a replication mode, not a tool. Airbyte and Fivetran both support CDC for Postgres, MySQL, MongoDB, SQL Server. Under the hood it's logical replication slots, binlogs, or oplogs — the connector reads the database's write-ahead log and emits events. The interview-relevant gotcha is that CDC slots hold WAL on the source until consumed, so a stuck connector can fill up your primary's disk. Always alert on pg_replication_slots.confirmed_flush_lsn lag.

Does dlt replace Airbyte and Fivetran?

Not yet, but for engineering-heavy teams it's a serious alternative. dlt's pitch is the pipeline is code — versioned, tested, deployed like any Python project. The trade-off is that you write more of it yourself. Teams that already have strong Python and orchestration practice (Airflow, Dagster, Prefect) often prefer dlt because it composes cleanly with their existing stack. Teams without that muscle should start with Airbyte or Fivetran and revisit later.

What's a realistic monthly EL spend for a Series A startup?

For a typical Series A pulling Stripe, Salesforce, HubSpot, Mixpanel, and Postgres into Snowflake, expect $1.5k-$6k/month on a managed tool at low MAR. Self-hosted Airbyte cuts the license cost to zero but adds 0.3-0.5 FTE of platform work, which at Series A salaries is $60k-$100k/year — a wash with the managed bill. The decision should pivot on whether that engineer's time is better spent on EL infrastructure or on modeling, governance, and product analytics. Usually it's the latter, which is why most early-stage teams start managed and migrate later.

How do I prepare for this question in interviews?

Pick one stack you've actually used and be ready to defend the choice with numbers: how many sources, how many MAR, what the bill was, what broke at 2am. Then learn the other stack well enough to compare honestly — read the Airbyte CDK docs and the Fivetran connector list, watch one talk from each company's annual conference, and play with the free tiers. The interview gold isn't I know all the tools; it's I picked this one, here's why, here's what I'd change.