Running on Neon?

Neon Postgres branching or suspending at bad times? We sort it out.

A preview environment spun up a branch for every pull request six months ago and nobody is tracking what came of them, the app's first request after a quiet period waits two seconds for compute to warm, and someone shipped a migration against main without a branch to try it on. We clean the project up and leave it in a shape you can hand to the next engineer.

  • Since 2015Fixing technology
  • FreeEstimates
  • US teamReal engineers

neon.log

still broken

What it is

Neon at a glance

Neon is serverless Postgres: a managed PostgreSQL service with storage separated from compute, so compute endpoints can scale down to zero when idle and scale back up when a connection arrives. The company started in 2021, built a copy-on-write storage engine on top of S3-compatible object storage, and introduced Git-like branches so a developer can create an isolated copy of production in seconds for testing a migration or a pull request. Databricks acquired Neon in 2025 and has continued to run it as a standalone product.

The pricing model has a free tier, included compute hours and storage on paid plans, and overage charges for compute, storage and egress above the plan limits. The two things that catch teams off guard are the compute auto-suspend, which produces cold starts after idle periods, and the branch limit on each plan, which is easy to hit if preview environments or CI jobs create a branch per pull request and never clean them up.

Type
Serverless managed PostgreSQL
First released
2022 public beta
License
Proprietary SaaS, open-source engine
Maintained by
Neon, Inc., a Databricks company
Typically used for
Web and mobile backends, preview environments, multi-tenant SaaS, AI apps paired with pgvector

The honest version

Neon pros and cons

No technology is all upside. This is where Neon earns its place and where it tends to cause trouble.

Where Neon is strong

  • Branches that behave like GitA branch is a point-in-time copy of production that is ready in seconds, without copying data. Preview environments and migration rehearsals stop being a scheduling exercise.
  • Compute scales to zeroIdle projects cost almost nothing, which is why Neon is so popular with multi-tenant SaaS that keeps a database per customer.
  • Standard Postgres on topIt is real Postgres: pg_dump, logical replication, extensions such as pgvector and PostGIS, and the same SQL your team already writes.
  • Point-in-time restore is first classYou can restore any branch to any second within the retention window, which is the recovery story most teams want but never get around to building.
  • Good fit with modern frameworksThe serverless driver, WebSocket proxy and HTTP endpoint make it comfortable from Vercel, Cloudflare Workers and other edge runtimes that struggle with classic TCP connections.
  • pgvector is supported out of the boxTeams building AI features can use the same database for application data and embeddings, which keeps the architecture small.

Where Neon bites

  • Cold starts are realA compute endpoint that scaled to zero has to resume on the next request, which adds latency measured in seconds rather than milliseconds. Low-traffic apps feel this more than they expect.
  • Branch limits hit quickly in CIA branch per pull request plus a branch per developer plus some orphaned ones from six months ago, and suddenly new branches fail to create.
  • Compute and storage overages add upIncluded compute hours sound generous until an app with constant traffic uses them in a week. The pricing is clear, but nobody watches the meter by default.
  • Pooling still needs attentionServerless does not remove the classic Postgres connection problem. Serverless drivers help but a connection pool in front of the compute is still the right answer at any real scale.
  • Some extensions and settings are not exposedShared libraries, superuser options and a few community extensions are not available because of the hosted model, which rules out certain workloads.
  • Ownership question after the acquisitionNeon still runs as its own product under Databricks, but roadmaps and pricing could shift, which is worth noting for teams making a long bet on it.

Sound familiar?

Common Neon problems we fix

The Neon problems that bring people to us, and what you have once each one is fixed.

too many branches: limit reached for project

Branch limit is blocking new work

Preview environments or ephemeral CI jobs have been creating branches without deleting them, and the plan's branch quota is full. New branches cannot be created until the old ones are removed, which stops pull-request previews from spinning up.

Branches lifecycled, limit no longer in the way

compute endpoint suspended, resuming

First requests are slow after idle periods

Compute scaled to zero after the auto-suspend window, and the next request has to wait for the endpoint to resume. For low-traffic or background-job workloads this happens all day.

Suspend configured to match actual usage

Compute hours used: 1,284 of 300 included

The compute bill ran over the plan

An app with steady traffic or a service that keeps connections open burned through the included hours well before the month ended, and overage charges are accumulating.

Right plan or right suspend, bill explained

FATAL: password authentication failed

Connections fail from the application

A role whose password was rotated in the dashboard but not in the deploy, or a serverless function with a stale secret, or a connection string missing the required endpoint ID.

Credentials rotated and in sync

Migration failed on branch

A migration broke on main before anyone tried it on a branch

The team skipped the step Neon exists to make easy. A branch would have shown the failure before it reached production.

Branch-first migrations, with rollback

Database egress: 420 GB this month

Egress charges are climbing

An application running outside the same region as the Neon project, a large analytics job scanning full tables, or a logical replication stream mirroring to a warehouse without pooling.

Traffic shaped, egress inside budget

Branch last restore point: 14 days ago

Point-in-time recovery is shorter than it looks

Retention on the current plan is shorter than people assume, and restoring to "a week ago Tuesday" is not always possible without upgrading first.

Retention set against the real recovery target

WebSocket connection closed unexpectedly

Edge runtime connections drop under load

A serverless or edge function reusing connections more aggressively than the proxy expects, or hitting the connection limit on a small compute size.

Driver, pooling and compute size matched

What we do

Neon help, from branch cleanup to ongoing support

Call us for one cold-start problem, or let us tidy the whole Neon project.

Branch cleanup and CI
Preview environments and CI branches that automatically age out, with the right mix of ephemeral and long-lived branches so the limit stops being a blocker.
Performance
Compute sizing, connection pooling, suspend behavior and indexes reviewed against your workload, with latency and bill measured before and after.
Migrations
Moving onto Neon from a classic managed Postgres, or from Neon to another host, done with real rehearsal on a branch rather than a crossed-fingers deploy.
pgvector and AI apps
Embedding tables, index types, dimension and ANN parameters set against your retrieval quality target, inside the same Neon project as your application data.
Backup and recovery
Point-in-time retention set to a sensible window, external pg_dump copies for the paranoid, and a tested restore you have actually run.
Managed Tech
Someone watching compute hours, branch counts and egress every week, with alerts before overages rather than after, and a team on call when a deploy breaks a connection string.

How it works

From first message to fixed

  1. Look Free estimate

    Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.

  2. Fix

    We fix the cause, not the symptom. If something is down, getting it back up comes first.

  3. Harden

    Backups, monitoring, updates and access, so the same problem can't quietly come back.

  4. Hand back or stay on

    You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.

Often found together

Running Neon alongside something else?

We work on the whole stack, so the problem can be wherever it actually is.

Questions

Neon questions

Something else? Call+1 737 727 7327

Why does our Neon database feel slow on the first request?

Compute endpoints suspend after an idle period to save money, and the first request after suspend has to wait for the endpoint to resume. That cold start is typically a second or two. For user-facing traffic this is often unacceptable, and the fix is either a longer auto-suspend window, a small keep-alive, or moving a latency-critical workload to a compute that never suspends.

How do we stop hitting the branch limit in CI?

Treat branches like preview deployments: create one per pull request, tear it down when the pull request closes or merges, and run a daily cleanup for anything orphaned. Separate long-lived branches such as staging from ephemeral ones, and only count the long-lived ones against your plan's headroom. We set this up through the Neon API or a CI script that runs on PR events.

Can you take over a Neon project a previous developer built?

Yes. We review the branches, roles, connection strings, extensions and compute sizes, move the project into your own organization, rotate every credential, and write down what each database is actually used for. Then we decide together which branches stay and which get deleted.

Is Neon a good fit for AI applications with pgvector?

For many teams, yes. pgvector is available, branches make it cheap to try a new embedding model against a copy of production, and keeping embeddings and application data in one database simplifies the architecture. For very large vector workloads a dedicated vector database may still be the right call, and we will model both before recommending one.

Should we move from a classic managed Postgres to Neon?

It depends on how the application uses the database. Branching, point-in-time restore and scale-to-zero are strong reasons to move. Steady high-traffic workloads with latency budgets measured in milliseconds, or workloads needing extensions Neon does not expose, are reasons to think twice. We will cost both based on your actual traffic rather than on a marketing page.

How much does Neon help cost?

Our estimates are free. Share the project or a description of what is happening, and we come back with a price and a short plan before any changes are made.

Free estimate

Neon Postgres giving you trouble?

Tell us what you're running and what's going wrong. A person reads it and comes back with an estimate.

  • No charge to find out what it will take
  • If it's urgent, say so

Replies come from a person. By sending, you authorize us to email you back. No marketing.