37signals · DHH
$10M / 5 yr
Moved seven apps including HEY off AWS onto owned hardware. Savings passed $1M in year one and ran ahead of projections.
Source [5] →SOLUTIONS · CLOUD EXIT
Stop renting your database
Cloud exit isn't sentiment — it's arithmetic. This page is the ledger: unit prices, receipts, and the road out, every figure with a source.
Managed databases solved a real problem: running PostgreSQL well used to take a specialist, and renting one was the rational call. But year by year the DBA's craft — HA, backup and recovery, monitoring — has been built into open-source software. More and more, the rent buys scarcity that no longer exists.
Meanwhile the price curves have split for good. One vCPU running for one month: about ¥19 on a self-hosted dedicated server, ¥270–432 on Aliyun's HA RDS, ¥580–1,582 on AWS[1]. Hardware gets cheaper every year; RDS list prices barely move. Every year you stay, the gap between renting and owning compounds.
Owning doesn't mean going it alone. Pigsty packages the whole enterprise-PostgreSQL stack as open source, with the author's subscriptions and expert services behind it. Below: the ledger, then the road.
THE LEDGER
The three big line items, priced piece by piece
Database, compute, and object storage — reduced to common units, rent laid next to own. The stubby self-host bars are the whole argument.
01 · DATABASE · ¥/vCPU·MO
- IDC self-host (Pigsty) ¥19
- Aliyun ECS (1:8 RAM) ¥180
- Aliyun RDS HA (1:2) ¥270
- Aliyun PolarDB Enterprise (1:8) ¥400
- Aliyun RDS HA (1:8) ¥432
- AWS RDS single-AZ (1:4) ¥580
- AWS RDS Multi-AZ (1:4) ¥1,160
- AWS RDS Multi-AZ (1:8) ¥1,582
02 · RAW COMPUTE · ¥/vCPU·MO
- IDC container (5× oversold) ¥7
- IDC dedicated server ¥19
- UCloud VM (oversold) ¥25
- AWS c5d.metal, 3-yr prepaid ¥80
- Aliyun ECS (1:2, dedicated) ¥107
- Aliyun ECS (1:8, dedicated) ¥180
- AWS EC2 M5, on-demand ¥301
03 · OBJECT STORAGE · ¥/TB·MO
- AWS S3 Express One Zone ¥1,120
- AWS S3 Standard ¥161
- Self-host MinIO/Silo (3-replica) ¥44
- Bare enterprise NVMe (baseline) ¥15
AWS S3 Standard
Silo/MinIO self-host ×3
Bare NVMe baseline
View as table
| Option | 1 TB | 10 TB | 100 TB | 1 PB |
|---|---|---|---|---|
| AWS S3 Standard | ¥161/mo | ¥1,610/mo | ¥16,100/mo | ¥164,864/mo |
| Silo/MinIO self-host ×3 | ¥44/mo | ¥435/mo | ¥4,350/mo | ¥44,544/mo |
| Bare NVMe baseline | ¥15/mo | ¥145/mo | ¥1,450/mo | ¥14,848/mo |
CALCULATOR
Price out your own cluster
Set eight controls — vCPU, RAM ratio, nodes, storage, billing term, subscription tier, currency, and monthly or annual view — to see what renting versus owning really costs.
RAM ratio
Term
Currency
Show
Aliyun RDS
AWS RDS
Aliyun ECS + Pigsty
IDC self-host + Pigsty
AWS has no standard RDS class at a 1:2 RAM ratio — that row is hidden for this selection.
View cost breakdown
| Option | Compute | Storage | Subscription | Total | Multiple |
|---|
Model: total = compute + storage (+ subscription on self-host rows). RDS at 1 node = basic single instance; 2 = HA pair; more = HA pair + read replicas (estimated). AWS long terms use 3-yr reserved (no 5-yr tier). Storage priced at Aliyun ESSD PL1 ¥1/GB·mo, AWS gp3 $0.08/GB·mo, self-host NVMe ¥14.5/TB·mo; cloud storage is counted once, conservatively (real bills add replica storage), self-host carries one full copy per node. Discount ladder: on-demand 1.5× · monthly 1× · 3-yr 0.45× · 5-yr 0.32× (Aliyun's ladder, applied to ECS as an estimate); no term discount applied to storage. People costs excluded on both sides.
PRICE TAGS
For reference: what everyone else charges
| Product | List price | Billing basis |
|---|---|---|
| Oracle Enterprise + RAC | ≈ ¥28,000 | / vCPU · year (22% annual-fee basis) |
| EDB BigAnimal Enterprise | $183.3 | / vCPU · month |
| Fujitsu PostgreSQL for K8s | $3,200 | / core · year |
| AWS RDS for PostgreSQL (HA) | ¥1,160–1,582 | / vCPU · month |
| Aliyun RDS for PostgreSQL (HA) | ¥270–432 | / vCPU · month |
| Pigsty subscription | $8K–60K | / node · year — vCPU uncounted |
"A fair price for decent database professional service is ¥10–20K a year." — commercial databases bill per core, so bigger boxes cost more; Pigsty bills per node and doesn't count your cores. Basis in the sources[svc].
METHODOLOGY & SOURCES
- Database & compute reduce to ¥ per vCPU·month (monthly instance price ÷ vCPU); object storage to ¥ per TB·month. Aliyun = east-China list prices (captured 2025-04); AWS = us-east at 7.24 CNY/USD; disk prices and subscription terms captured 2026-08-06.
- The ¥19/vCPU·mo self-host figure is a dedicated 64C/384G server over five years — same order as Tantan's (¥19.5) and DHH's (~¥25) published numbers; NVMe at ¥14.5/TB·mo is an enterprise drive over five years.
- Aliyun RDS single-instance ≈ HA price ÷ 2.35 (the list-price ratio across the dedicated basic series); AWS Multi-AZ = 2 × single-AZ (list relationship).
- Prices move — verify against the vendors' own pages: AWS RDS for PostgreSQL pricing · Alibaba Cloud price calculator.
- Are Cloud Databases an IQ Tax? — vonng.com/en/cloud/rds/
- Analyzing Alibaba Cloud Compute Cost — vonng.com/en/cloud/ecs/
- Is Cloud Block Storage a Slaughter Scam? — vonng.com/en/cloud/ebs/
- Paradigm Shift: From Cloud to Local-First — vonng.com/en/cloud/paradigm/
- DHH: Cloud-Exit Saved Over Ten Million — vonng.com/en/cloud/odyssey-done/
- How Ahrefs Saved $400M by Not Going Cloud — vonng.com/en/cloud/ahrefs-saving/
- Did RedNote Exit the Cloud? — vonng.com/en/cloud/rednote-cloud-exit/
- Dissecting Cloud Object Storage — vonng.com/en/cloud/s3/
- [svc] Pigsty subscription & service terms — https://pigsty.cc/docs/about/service/ (retrieved 2026-08-06)
- [ebsprice] AWS EBS pricing (us-east-1) — https://aws.amazon.com/ebs/pricing/ (retrieved 2026-08-06)
- [essdprice] Aliyun cloud-disk price reference — https://www.foreignserver.com/aliyun/sjpjiage2025.html (retrieved 2026-08-06)
RECEIPTS
The ones who left already showed their math
Ahrefs
$400M / 3 yr
The search-index giant priced its own datacenter against AWS list prices, line by line: staying off the cloud saved roughly $400M over three years.
Source [6] →Tantan · 探探
25,000 vCPU
Runs PostgreSQL & Redis on ~25,000 self-hosted cores at millions of QPS, with a DBA team of two to three — the ¥19.5 core·month reference behind the chart above.
Source [1] →RedNote · 小红书
Self-host first
A platform that grew up in the cloud-native era, now steering its infrastructure strategy toward self-hosting first.
Source [7] →These are public, third-party stories. Our own customers' names appear here only with written permission.
THE ROAD
Four steps out
Each step maps to a service tier that can catch you — from a single consultation to full coverage, your call.
1Assess
- Inventory instances, specs, dependencies
- Run the TCO math on your real bill
- Flag the hard parts, order the moves
≈ Quick Consultation $30 / issue · Expert Session $400 / session
2Pilot
- One spare machine, one command
- Move a non-critical workload first
- Prove HA & PITR on the dashboards
≈ Open Source · Free
3Migrate
- Logical replication keeps data in sync
- Rehearse cutover and rollback
- Execute in a planned, minutes-long window
≈ Emergency Response $4,000 / day
4Operate
- Self-healing HA & automated backups take the routine
- Dashboards and alerts take the night shift
- A subscription takes the hard cases
≈ Subscription from $8K/year
OBJECTIONS
The three pushbacks you'll hear
"Self-hosting is an ops burden."
It was, when HA, backups, and monitoring were hand-built. Pigsty ships them as software — it gives a junior engineer 70%+ of a professional DBA's capability — and the author's expert services stand behind the hardest stretch. Tantan's two-to-three-person DBA team running 25,000 cores is this logic, proven.
"We can't do HA as well as the cloud."
You don't have to build it: self-healing failover on Patroni + etcd and point-in-time recovery with pgBackRest ship as defaults. A real cluster runs in the open at demo.pigsty.io — go look.
"Compliance won't allow it."
Self-hosting is usually the compliant path, not the risky one: data stays in your jurisdiction on your keys, offline and air-gapped delivery is supported, and the Ansible IaC leaves an audit trail. Contact us to review the evidence and boundaries for your environment.
THE ESSAYS
Read the whole argument
This page is the digest. The full case — data, receipts, and philosophy — lives in the author's four-year cloud-exit column. These essays are the way in.
The math
The receipts
Let's price out your exit
Send a recent invoice and a spec list — you'll get a straight answer on whether the math works for you.
FAQ
Cloud-exit questions, answered
Is self-hosting really cheaper than cloud RDS?
For steady production workloads, yes — usually by multiples. The numbers above are list prices you can verify; run your own spec through the calculator, then check it against a real invoice. The gap widens with scale.
Don't we need a full-time DBA to self-host?
Pigsty packages top-tier DBA judgment into software — it gives a junior engineer 70%+ of a professional DBA's capability. For the hard remainder, subscriptions and expert services are served by the people who wrote the code.
What about high availability and disaster recovery?
Self-healing HA on Patroni and etcd, point-in-time recovery with pgBackRest, and 600+ metrics per instance watching it all — the guarantees RDS sells, in software you control. A live cluster runs in the open at demo.pigsty.io.
How disruptive is the migration itself?
The path is incremental: assess first, pilot on a spare machine (one command), then cut over with logical replication and a rehearsed rollback plan. When logical replication fits your schema, the cutover window comes down to minutes — and the assessment tells you whether it fits.
We're not even on PostgreSQL yet — does this still apply?
Yes. Pigsty ships 12 swappable kernels with MSSQL / Oracle / MySQL / Mongo compatibility options, and Oracle-to-PostgreSQL migration is one of our standard expert engagements.