blog
Blog — page 23 of 39
Posts 529–552 of 928, newest first.
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.
Ajay KumarMocks 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.
Ajay KumarSharding 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.
Ajay KumarWiring 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.
Ajay KumarKeeping 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.
Ajay KumarDebugging 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.
Ajay KumarHow 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.
Ajay KumarSandboxing 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.
Ajay KumarPer-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.
Ajay KumarMicroVM 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.
Ajay KumarIsolating 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.
Ajay KumarPandaStack 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.
Ajay KumarPandaStack 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.
Ajay KumarRunning 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.
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.
Ajay KumarE2B 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.
Ajay KumarBest 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.
Ajay KumarSandboxing 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.
Ajay KumarAI 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.
Ajay KumarAI 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."
Ajay KumarFirecracker 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.
Ajay KumarPandaStack 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.
Ajay KumarBest 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.
Ajay KumarBest 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.
Ajay Kumar