blog
Blog — page 27 of 39
Posts 625–648 of 928, newest first.
Per-Tenant AI Voice Transcription in MicroVMs
Your transcription service holds a customer's support calls, a clinic's dictation, and somebody's deposition in the same Python process, with ffmpeg parsing whatever container format arrived. A malformed .wav is an unsolicited code-execution proposal. One microVM per tenant makes it a boring one.
Ajay KumarRunning Per-Tenant Billing and Usage Metering in MicroVMs
Most isolation bugs leak data. This one mails a customer an invoice built partly from a competitor's usage. The only bug class where the incident review includes your CFO — so make cross-tenant reads structurally impossible, not merely unlikely.
Ajay KumarTesting Browser Extensions with AI Agents in MicroVMs
An extension gets content-script access to every page you visit, your cookie jar, and a background worker that can phone home. Installing one to 'just test it' on your laptop is a trust decision. Do it in a microVM you throw away.
Ajay KumarHow to Sandbox an Untrusted composer install
composer install is a build step the way a stranger's USB stick is a file transfer. Scripts and plugins run arbitrary PHP before your app loads a single class — here's how to contain that.
Firecracker vs Amazon ECS Anywhere: Two Different Questions
ECS Anywhere answers "can AWS schedule things on my hardware?" Firecracker answers "can each workload get its own kernel?" They're not competitors — they're answers to different questions, and the crux is that external ECS instances run containers on your shared host kernel, not Fargate's microVMs.
Ajay KumarFirecracker virtio-net RX/TX Queues Explained
A virtqueue is a ring buffer with commitment issues. Here's how a packet actually crosses the guest/host boundary in Firecracker, why there's one RX/TX pair per net device and no multi-queue, and the tuning levers that really exist — plus how to measure instead of guess.
Ajay KumarNUMA Locality and Firecracker Snapshot Restore
Your memory is local until the scheduler has an opinion. On a two-socket host, first-touch allocation means whichever thread faults a page decides which NUMA node it lives on for the rest of the VM's life — and with a userfaultfd handler, that thread isn't the vCPU. Here's the primer, the tooling, the mitigations, and the honest case for not building any of it.
Ajay KumarBest Sandbox APIs for Python Coding Agents in 2026
You're a Python developer wiring a sandbox into LangGraph, CrewAI, or a plain Anthropic tool-use loop. This one is about the SDK ergonomics you'll actually live in — pip install to first exec, streaming, statefulness, cancellation — plus an honest pass over the field.
Ajay KumarSnapshot Restore vs Process Preforking: Same Idea, Different Boundary
Unicorn, gunicorn, uWSGI, Android's Zygote, Chrome's zygote — everyone reinvented fork() to stop paying initialization twice. It works beautifully, until you notice your workers share one kernel, one user, and one network stack.
Ajay KumarPandaStack vs Koyeb: an honest comparison
Koyeb deploys and scales your services globally from a repo or image. PandaStack creates disposable, forkable Firecracker microVMs for untrusted, agent-generated code. Where they overlap, and where they don't.
Ajay KumarSnapshot Restore vs Cold Boot: The Tradeoffs Nobody Lists
A restored VM is a machine that skipped the last two months and is extremely confident about it. Here's the full bill for not booting: frozen sizing, a lying clock, cloned entropy, and secrets baked into a file you replicate to object storage.
Ajay KumarSandboxing User-Written Webhook Transformations
Somebody added a textarea labeled 'Transform (optional)' and shipped it on a Thursday. Congratulations: you are a code-execution company now, and nobody told your threat model.
Ajay KumarPandaStack vs AWS Lambda for Running Untrusted / LLM-Generated Code
AWS invented Firecracker, so the isolation argument is over before it starts. The real question is whether a function-shaped deployment unit fits code an LLM wrote thirty seconds ago.
Ajay KumarThe Best AI Code Execution Platforms in 2026
Your agent will eventually generate a confident rm -rf. Where that runs is the most consequential infrastructure decision in the whole stack — here's how to choose, from someone who built one of the options and will tell you when to pick a different one.
Ajay KumarGuest Clocks, Snapshots, and the Time Travel Problem
A restored microVM is a machine that is confidently living in the past. It doesn't know it. Your TLS stack finds out first, and gets shot for being the messenger.
Ajay KumarReplicating Firecracker Snapshots Across Regions
A snapshot is not a file you can cp. It's a set of artifacts that only mean something together, and shipping that set to another region is a distributed publish problem wearing a storage costume.
Firecracker vs Bottlerocket: One Is a Hypervisor, One Is a Host OS
These two aren't rivals; they're stacked. Bottlerocket hardens the host you share. Firecracker removes the sharing. If you're running untrusted code, that distinction is the whole ballgame.
Ajay KumarPer-Tenant Backup and Restore, Isolated by a MicroVM
A backup worker can read everything by design and a restore worker writes bytes a customer chose. Put a hypervisor boundary and a scoped credential around each job, one tenant at a time.
Ajay KumarRunning a WebAssembly Plugin Host Inside a MicroVM
Wasm sandboxes the plugin from your process. It does not sandbox the plugin from the host function you helpfully named exec_sql.
Ajay KumarPer-Tenant Isolation for AI Training-Data Labeling Pipelines
Filtering by tenant_id protects you from bugs in your query, not from bugs in your process. And once one tenant's documents land in another tenant's training set, the model has already eaten it — you cannot un-train a leak.
Ajay KumarTimeouts and Cancellation for AI Agent Tool Calls
The model wrote while True: pass and it will do that with total sincerity, forever, on your bill. In-process timeouts don't stop it. A destroyed microVM does.
Snapshot Restore vs Container Image Pull: Two Ways to Start Fast
Your container started in 200ms and then spent nine seconds importing pandas, which is a kind of performance. Lazy pulling attacks the bytes; snapshot restore attacks the part that actually costs you — a process that has already finished initializing.
Ajay KumarHow Sandbox Pricing Models Actually Work in 2026
Nobody's bill is decided by the headline rate. It's decided by whether you pay for the seconds your agent spends watching an LLM think. A vendor-neutral guide to the pricing axes.
Ajay KumarThe Snapshot Clone Randomness Problem
Fork a snapshot fifty ways and you get fifty microVMs that agree on what a random number is — the one thing they should never agree on. Here's exactly what freezes, why duplicate UUIDs and reused ECDSA nonces are the real risk, and which layers VMGenID does and does not repair.
Ajay Kumar