Your next tar backup to S3 just got easier. Or does it?
Developers everywhere groan when S3 forces object semantics on file workflows. No symlinks. No appends without hacks. AWS S3 Files — launched April 7 — vows to fix that. Mount your bucket. Use cat, ls, Python’s open(). Act like it’s a real filesystem. Sounds dreamy. But excitement drowned out the tests. Almost nobody ran the numbers.
So I did. Throwaway account. t4g.micro instance. Real ops. Here’s the missing conversation.
Why the AWS Hype Machine Ignores the Cold Cache Trap?
Andrew Warfield, AWS VP, sold the vision clean: > “With Tables, Vectors, and now Files, we are consciously changing the surface of S3. It’s not just objects — it’s evolving to make sure you can work with your data however you need to.”
Noble. S3 as universal data layer. Tables for queries. Vectors for AI. Files for POSIX. But paper plans crash on metal.
It’s NFS 4.1/4.2 over S3. Powered by EFS guts. Data lives in S3. Hot files cache in fast storage. Mount helper? Buried in amazon-efs-utils. IAM? elasticfilesystem.amazonaws.com. Mount shows as nfs4 on localhost. EFS veterans: zero learning curve. Newbies: brace for security groups, access points, ClientWrite policies. Launch blog glosses that over.
Performance? Headline screams ‘terabytes per second.’ My baseline — single AZ, tiny instance — begs to differ. First read of 5MB file: 17 MB/s. Yawn. Second read: 3.1 GB/s. Whoa. Cache warmed up. Small files? 2-6 ms. Writes flew.
But — and here’s the kicker — it hinges on that 128KB threshold. Tiny files cache hot. Big ones stream from S3. Cold access? S3 lag with NFS overhead. Your ML training set scattered across cold blobs? Pray for repeats. One-offs? Stick to SDKs.
Is AWS S3 Files Production-Ready or Just a Fancy Proxy?
Tested real ops. Here’s the scorecard:
| Operation | Result |
|---|---|
| Write, append | Works |
| Rename (mv) | Works — atomic |
| mkdir -p | Works |
| chmod | Works — POSIX preserved |
| Symlinks | Works (sorta) |
| File locking (flock) | Works |
| Hard links | Nope — “Too many links” |
Read-after-write? Consistent. No stale nonsense. Symlinks sync as objects holding paths — not pure symlinks. Hidden .s3files-lost+found dir lurks. Bidirectional sync keeps bucket fresh.
Luc van Donkersgoed nailed it: > “Does this finally settle the question ‘is it a file or is it an object’?”
Yes and no. It’s both. Mount as files. Query as objects. Genius? Or identity crisis?
Look, this isn’t revolution. It’s evolution — smart, but incremental. EFS loyalists won’t ditch; S3 Files lacks multi-AZ mounts out-gate (coming soon, AWS whispers). Costs? Metered like EFS — throughput, IOPS. Your exabytes idle? Bills stack.
My unique take: This echoes the 90s NFS explosion. Back then, NFS turned LANs into shared filesystems, killing sneaker nets. S3 Files does that for cloud — but only if your pattern fits the cache. Miss? You’re back to rclone pain. Prediction: 80% of users love it for backups, CI/CD. The rest? Hype-chasers bolt.
AWS PR spins ‘universal substrate.’ Cute. Reality: Niche killer for POSIX-on-S3. If you’re gluing legacy tools to S3 — bioinformatics pipelines, video editors — jackpot. Everyone else? Meh upgrade.
And the EFS ties? Sneaky. Forces EFS knowledge down S3 throats. Unified? Or vendor lock-in 2.0? AWS doesn’t care; you’re in their zoo.
Short para. Brutal truth.
Scale it up. Multi-instance? Shared cache? Docs vague. Throughput caps? Per-mount? Test farms pending.
But damn, warm cache feels like NVMe. 3 GB/s on micro? Scale to m7g.mental? Terabytes await.
Corporate spin calls it ‘platform shift.’ I call bullshit on untempered joy. Test first. Or regret.
What Does AWS S3 Files Cost, and Is It Worth It?
Pricing mirrors EFS: storage in S3 (cheap), plus mount fees — throughput ($0.30/GB out), IOPS bursts. Cache fills from S3 GETs — surprise egress? Ouch for cross-region.
Worth it? If S3’s your lake and NFS your drug — yes. Otherwise, EFS direct. Or MinIO on EC2 for open-source purity (we’re Open Source Beat, after all).
Dry humor: AWS finally admits objects suck for files. Took ‘em 17 years.
Wander here — remember Glacier restores? This caches faster, but cold still bites.
Unique insight: Parallels Hadoop’s HDFS — promised POSIX, delivered wonky. S3 Files nails consistency better. Won’t repeat that mess.
🧬 Related Insights
- Read more: GPU Buying Guide 2025: Cards That Fuel Tomorrow’s AI Dreams on Any Budget
- Read more: 93 Gigawatts of AI Inference Compute by 2030: Kubernetes Steps Up to Standardize It All
Frequently Asked Questions
What is AWS S3 Files?
Mounts S3 buckets as NFS filesystems. POSIX ops on object data. Caches hot files.
How fast is AWS S3 Files in real tests?
Cold: 17 MB/s. Warm: 3+ GB/s. Small files: ms latencies. Pattern-dependent.
Does AWS S3 Files replace EFS?
No. S3-backed, EFS-powered. Cheaper storage, same ops complexity. Pick by workload.
