Fivetran: Expensive Data Plumbing That Actually Works

Fivetran is what happens when engineers get tired of fixing data pipelines every weekend. Started in 2012, now it's where you end up after your third "the pipeline is down" call from your boss.

How This Thing Actually Works

Fivetran uses ELT instead of ETL, which means it dumps raw data first, then you transform it later. This sounds backwards until you realize it's genius - no more losing data because some transformation fucked up halfway through.

Here's what you get:

• 1-minute syncs if you pay for enterprise (most people don't need this)
• Schema changes don't break everything - it just adds columns automatically
• Handles huge amounts of data - scales to petabytes if your wallet can handle it
• Actually works most of the time, unlike that Kafka Connect setup that keeps dying with ERROR Failed to flush WorkerSinkTask{id=salesforce-sink-0} because someone thought 16 partitions for customer data was a good idea

The platform comes in three flavors: SaaS (easiest), hybrid (for paranoid security teams), and private cloud (for really paranoid security teams).

The Connector Situation

They claim 700+ connectors but here's the reality:

• Salesforce, HubSpot, Google Ads: Work great, rarely break
• Weird SaaS tools: Hit or miss, good luck debugging when they fail
• Databases: PostgreSQL and MySQL work flawlessly; Oracle costs extra and is a pain
• Files: S3 works fine, SFTP makes you want to quit (why is it always permission denied?)
• Kafka: Why are you using Fivetran for this?

The Good Parts

It works most of the time: I've had one Fivetran connector randomly die in two years - PostgreSQL connector just stopped mid-sync with zero useful logs, just "Connection terminated unexpectedly" in CloudWatch. Took 3 hours on support chat to figure out our RDS instance hit connection limits because some asshole left a connection pool running in a Lambda. Compare that to the Airbyte setup that throws docker: Error response from daemon: OCI runtime create failed every time the wind changes direction.

Schema changes don't kill you: Remember spending a weekend fixing your pipeline because Marketing added a field to HubSpot? Fivetran just handles it. New columns appear, data keeps flowing.

Security nerds love it: SOC 2, ISO 27001, all that compliance bullshit your security team demands. HIPAA, GDPR, PCI DSS - they've got the certificates to prove it.

The Reality Check

When It's Worth the Money

• Your team spends more time fixing data pipelines than building analytics
• You've got enterprise budget and need reliability
• Compliance requirements make custom solutions a nightmare
• You're tired of being woken up by "data sync failed" alerts

When It'll Bankrupt You

• Startup with limited runway
• High-volume, low-value data (think IoT sensor spam)
• Custom data sources that need heavy transformation
• You actually enjoy writing ETL code (seek help)

Stuff That'll Go Wrong

The Bill Shock: Had a team where Salesforce sync went completely insane - their bill jumped from like $800 to something stupid, maybe $7K? Could've been $8K. Historical data sync is a budget killer.

The API Change: Zendesk deprecated their API v2 endpoints in July 2025 with zero fucking notice, broke our connector for three days straight. Fivetran finally pushed a fix on day 4, but we were completely dark on customer support data during our biggest sales week. Sales team kept asking "why are the conversion dashboards showing zeros?" Fun explaining that one to the CEO.

The Schema Thing: Column names changed and broke our dashboards. "Automatic" schema detection isn't smart enough to know that "customer_name" and "customerName" are the same thing.

Most teams using Fivetran are ex-ETL engineers who got tired of explaining why data was 6 hours late. It works, it's expensive, and you'll sleep better at night.

The Monthly Active Rows Mindfuck

Here's the deal: Fivetran pricing is based on "Monthly Active Rows" (MAR), which sounds simple until you realize it's designed to be impossible to predict. Every row that gets touched in a month counts, and they changed the model in March 2025 to be per-connector instead of account-wide. Translation: your bill just got more expensive.

What You'll Actually Pay

Free Plan: 500K MAR per connector. Sounds generous until you connect Salesforce and burn through it in 6 hours because every fucking opportunity update counts as a row change. Did a test sync on Friday, got a bill alert by Monday morning.

Standard Plan: Starts around $120/month, quickly becomes $500-2000/month once you hit real data volumes. The 15-minute sync limit means you're not getting real-time data despite what the sales team implied.

Enterprise Plan: $2000-10,000/month territory. You get 1-minute syncs, which is still not real-time but close enough for most use cases. This is where most companies end up after the sticker shock.

Business Critical: If you need to ask, you can't afford it. Think $20K+/month for the privilege of managing your own encryption keys.

Pricing Horror Stories From the Trenches

The Salesforce Thing: Team at my last job enabled historical sync on a 8-year-old Salesforce org with 14M opportunity records. Bill went from $240/month to $8,400 in one month because every record update since 2017 counted. Finance director lost his shit when the Amex alert came through. Nobody bothered to read the fine print about "all historical data" actually meaning ALL historical data.

The Connector Multiplication: Used to pay $1000/month for account-wide usage. 2025 pricing changes moved to per-connector billing. Same data usage, now $3000/month across 8 connectors. Each connector gets its own MAR threshold instead of sharing.

The dbt Thing: Transformations looked cheap until the bill showed up. Extra $500 because dbt runs kept spawning more models than anyone expected. Turns out ref() dependencies cascade like dominoes.

Implementation Reality Check

The "Quick Setup" Lie

Sales says "10 minutes to get started." Here's what actually happens:

1. Day 1: OAuth setup takes 10 minutes (this part's true)
2. Week 1: Realize you need custom schemas, spend time mapping fields
3. Week 2: Security team freaks out about data residency, need hybrid deployment
4. Week 3: Finally get approval, discover half your connectors need enterprise plan
5. Month 2: First real sync, everything breaks because of API rate limits you never knew existed

Performance Is Actually Good

When it works, it fucking works:

• Handles huge data volumes if your wallet can survive it
• Parallel processing means multiple connectors sync simultaneously
• Auto-retry actually works, unlike your Kafka setup
• Sub-second latency for database CDC (if you pay for HVR connectors)

The "sub-millisecond latency" thing is marketing bullshit. That's database replication latency, not end-to-end sync time. Your Snowflake warehouse still needs 2-3 minutes to process the batch.

Infrastructure Requirements They Don't Mention

SaaS is easiest but your security team will hate you. All data flows through Fivetran's servers.

Hybrid deployment requires you to run their agents in your cloud. Sounds secure until you realize you're still trusting their code with your data, just in your VPC.

Network setup is a nightmare: VPN tunnels, firewall rules, IP whitelisting. That "no infrastructure" promise evaporates quickly.

The Vendor Lock-in Trap

Easy to Get In, Expensive to Get Out

Migration away from Fivetran means rebuilding every pipeline from scratch. They know this. The extensive connector library isn't just convenient - it's a trap.

Your data is in standard formats (great!), but the extraction logic, scheduling, error handling, and monitoring all need to be rebuilt. Budget 6-12 months of engineering time to migrate off.

What Actually Makes Teams Choose Fivetran

Exhaustion: Your team is tired of 3am pages about broken data pipelines
Compliance: SOC2, HIPAA, PCI DSS boxes are already checked
Reliability: It actually works 99% of the time (that 1% will still wake you up)
Speed: Get data flowing in days, not months thanks to pre-built connectors

When to Run Away

High-volume, low-value data: IoT sensors, application logs, anything with millions of rows of cheap data
Custom sources: If your data source isn't in their 700+ connectors, you're fucked
Tight budget: That $120/month quickly becomes $2000/month in production
Control freaks: If you need to modify extraction logic, build custom pipelines instead

Bottom Line for Engineers

Fivetran costs 3-5x more than alternatives but actually works. It's the AWS of data integration - expensive but reliable. If your time is worth more than the premium, and you've got enterprise budget, go for it. If you're counting pennies or need custom logic, build it yourself or use Airbyte.

The pricing model is designed to grow with your data, which means it's designed to get more expensive over time. Plan accordingly.