blog

Blog — page 23 of 39

Posts 529–552 of 928, newest first.

9 min read

Getting realistic test data into ephemeral databases

Your fixtures have three users and five orders. Production has two million orders and a customer whose name contains an apostrophe. The bugs live in the gap between those, and closing it is a data problem.

testingdatabasesephemeral-environmentsAjay Kumar
8 min read

Mocks that pass while production breaks

Every mock is a claim about how something behaves. The test suite verifies your code against the claim, not against reality — and the gap between those is where the interesting outages live.

testingarchitectureci-cdAjay Kumar
8 min read

Sharding a Playwright suite that has outgrown one machine

A Playwright suite grows until the feedback loop is too slow to use. Adding workers helps until the machine runs out of memory — after that you need shards, and shards need state that isn't shared.

testingbrowser-automationci-cdAjay Kumar
8 min read

Wiring preview environments into pull requests

Reading a diff tells you whether the code is reasonable. A URL tells you whether the feature works. Getting from one to the other is about a hundred lines of CI and one decision about databases.

ephemeral-environmentsci-cdapp-hostingAjay Kumar
8 min read

Keeping a browser logged in without storing passwords

Most browser automation spends its first thirty seconds logging in, and most of its failures happen there too. There are three ways to skip it, and they differ in how much they keep.

browser-automationsecuritysnapshotsAjay Kumar
8 min read

Debugging a browser failure you can't reproduce

'Works on my machine' is a real technical statement about browser tests. The CI environment differs in about eight specific ways, and each one produces a recognisable failure.

browser-automationci-cdtestingAjay Kumar
8 min read

How many browsers can one machine actually run?

Everyone discovers the browser concurrency limit the same way: by exceeding it. The number is set by memory, it's lower than you'd guess, and the failure past it looks like flaky tests.

browser-automationperformancemicrovmAjay Kumar
8 min read

Sandboxing AI Resume-Screening Agents With MicroVMs

A resume is a file you didn't write, from a person you haven't vetted, headed straight into an LLM prompt with the next candidate's data still warm in the same process. One microVM per candidate fixes that.

ai-agentsrecruitinghr-techAjay Kumar
9 min read

Per-Tenant MicroVM Isolation for Thumbnail Generation

A JPEG should not be able to read your AWS credentials. It can, though, if your thumbnailer shares a worker with every other tenant's uploads.

securitysandboxmicrovmAjay Kumar
9 min read

MicroVM Isolation for AI Invoice OCR and Extraction

An invoice-extraction agent is a program that runs an untrusted PDF parser, then an LLM that reads whatever text that parser found, then writes a number into your payment system. Give every invoice its own microVM.

invoice-ocraccounts-payableprompt-injectionAjay Kumar
9 min read

Isolating AI Translation Agents on microVMs

Your translation agent will render 'ignore all previous instructions' flawlessly into fourteen languages. That's not a bug in the model — it's a bug in your architecture. Fix it with per-tenant microVMs.

localizationtranslationai-agentsAjay Kumar
9 min read

PandaStack vs Neon: Comparing Managed Postgres Providers

Neon popularized instant Postgres branching through storage-compute separation. PandaStack takes a different path — a dedicated microVM per database on the same substrate as your compute. Here's the honest trade-off.

comparisonpostgresdatabasesAjay Kumar
9 min read

PandaStack vs Supabase: an honest comparison

Supabase gives your schema an API. PandaStack gives your code a machine. They solve adjacent problems from opposite directions, and most teams asking this question actually want to know which direction matches their problem.

comparisonsupabasedatabaseAjay Kumar
8 min read

Running dotnet restore and dotnet build on untrusted code in a microVM

MSBuild will happily execute a .targets file the moment dotnet restore touches a package — no dotnet run required. Here's how to make that safe to do on repos you don't trust.

securitydotnetsandboxAjay Kumar
10 min read

The best serverless function platforms in 2026

Every 'best serverless platform' list reads like a press release. This one compares isolation models, cold starts, and pricing shapes honestly — including where PandaStack does and doesn't fit.

serverlessfaasfunctionsAjay Kumar
8 min read

E2B vs Modal: Which One Actually Fits Your Workload?

E2B and Modal both show up in AI infra conversations, but one is a sandbox for running agent-generated code and the other is a serverless compute platform for Python/ML workloads. A fair head-to-head on where each actually fits.

comparisonai-sandboxe2bAjay Kumar
10 min read

Best Database-as-a-Service Platforms (2026)

Every team outsources database ops to somebody eventually — the question is which isolation model, backup story, and pricing curve you're signing up for. Here's an honest walk through the major managed-database options in 2026, including where PandaStack's dedicated-VM-per-database model fits and where it doesn't.

database-as-a-servicemanaged-postgrescomparisonAjay Kumar
9 min read

Sandboxing AI incident-response agents: isolating the runbook

An LLM that can read a P1 alert and run kubectl commands is either your best on-call engineer or the fastest way to turn a blip into an outage. The difference is where it's allowed to execute.

ai-agentsincident-responsesreAjay Kumar
9 min read

AI Contract Review Agents: Isolating Privileged Documents in MicroVMs

An LLM that hallucinates a clause is embarrassing. An LLM whose sandbox leaks one client's term sheet into another client's session is a malpractice claim with your firm's name on it.

ai-agentslegal-techmicrovmAjay Kumar
9 min read

AI agents that edit podcasts and video: sandboxing the render pipeline

The moment an editing agent runs ffmpeg on a file a stranger uploaded, your threat model stops being "untrusted code" and starts being "untrusted bytes owning the parser that reads them."

ai-agentsmedia-processingmicrovmAjay Kumar
9 min read

Firecracker vs crosvm: two rust-vmm siblings, two different jobs

They share a language, a KVM backend, and literal git history in rust-vmm. Past that, one was built to run thousands of headless server guests behind a REST API, and the other was built to put a graphical Linux desktop inside a Chromebook. The device model gives it away immediately.

firecrackercrosvmvmmAjay Kumar
9 min read

PandaStack vs Replit

Replit Agent builds an app for you, inside Replit's own browser tab. PandaStack doesn't build anything — it's the compute layer underneath somebody else's agent, including possibly yours. Different jobs wearing similar marketing copy.

comparisonreplitai-agentsAjay Kumar
10 min read

Best Postgres Database Branching Platforms (2026)

Every PR gets its own database, every AI agent gets a disposable one to run migrations against, and nobody argues over a shared staging box anymore — that's the promise of database branching. Here's an honest look at who builds it and how, so you can pick the isolation model that actually fits your team.

postgresdatabase-branchingcomparisonAjay Kumar
10 min read

Best Preview Environment Platforms (2026)

Every pull request can now get a live, clickable URL — reviewers want it, QA wants it, and increasingly AI coding agents need one to check their own work. A qualitative, hedged look at where Vercel, Netlify, Railway, Render, Northflank, GitHub Codespaces, and PandaStack each sit on isolation model, database handling, and boot latency.

preview-environmentsci-cdcomparisonAjay Kumar