blog
Blog — page 6 of 39
Posts 121–144 of 928, newest first.
virtiofs vs virtio-blk: How Files Actually Get Into a MicroVM
One gives the guest a disk it owns. The other gives it a window onto a directory the host owns. That single difference decides whether you can fork a machine in 400ms, who parses guest-controlled input, and what a multi-tenant escape looks like.
Ajay KumarWhat "persistent" actually means in a sandbox
You wrote the file. You ran cat and saw it. Neither of those facts says the bytes are on a disk. Here is every layer a write passes through inside a microVM, which of them a crash erases, and why the honest answer for an ephemeral rootfs is not "fsync harder" but "get the artifact out".
Ajay KumarInterrupts, IRQs, and Where microVM Tail Latency Comes From
Median latency tells you the machine works. p99 tells you how the machine is built. Here is every handoff a virtio interrupt makes inside a Firecracker guest, which of those handoffs are queues, and which ones you can actually do something about.
Ajay KumarWhy the disk under your sandbox fleet decides your boot time
People pick a sandbox platform on features and then get bitten by storage hardware. A reflink clone is metadata-only and nearly free; every copy-on-write byte afterwards is a real read-modify-write against a real device. Under 50 concurrent restores, that device is either local NVMe or it is your bottleneck.
Ajay KumarKnowing Which Customer Costs You Money
If you run code on behalf of customers, your cloud bill arrives as one number and your customers arrive as a list. Splitting the first across the second is a real engineering problem, and nearly everyone gets it wrong the same way: by attributing on wall-clock vCPU, which overcharges the bursty and undercharges the idle.
Ajay KumarCan You Run a Sandbox Fleet on Spot Instances?
The discount is real and the question people ask about it is the wrong one. Spot is not a pricing decision, it is a statement about whether your workload can be interrupted and rebuilt — which is a different answer for a 30-second code execution than for a customer's database.
Ajay KumarBare metal vs cloud VMs for running Firecracker
If you run microVMs, you run a hypervisor — and the layer underneath it is an architecture decision, not a procurement detail. Nested virt taxes every exit and every page fault; bare metal hands you the machine and the pager. Here is the ledger I actually use.
Ajay KumarA plain-English glossary of sandbox and microVM terms
This vocabulary is genuinely confusing, and a lot of it is vendors using one word for three things. Here is the glossary I wish someone had handed me: 48 terms, each with the distinction that actually matters and the specific way people get it wrong.
Ajay KumarYour benchmark ran at a different clock speed than production
The same core does not run at the same speed twice. Governor, turbo bin, how many neighbours are busy, thermal headroom and ramp latency all move it — which is why the first sandbox on a quiet host looks fast and the fiftieth on a busy one gets blamed on the platform.
Ajay KumarRate-Limiting the Network on a Per-Sandbox Basis
Everyone sets a CPU limit and a memory limit. Almost nobody sets a network limit, and then one tenant saturates the host NIC and every other sandbox on the box gets slow for reasons that appear in none of their own metrics. Here is where the enforcement point actually is, and which end of the veth pair it belongs on.
Ajay KumarSpectre, Meltdown, and What a microVM Actually Protects You From
KVM stops a guest from naming host memory. It does not stop the CPU from guessing about it. An honest tour of the transient-execution families for people running multi-tenant compute — how to read your host's actual exposure, what the mitigations really cost, and the part no VMM choice fixes.
Ajay KumarImmutable microVM Rootfs with dm-verity
"Read-only" is a promise the guest makes to itself. dm-verity is a Merkle tree the kernel checks on every block read — so tampering is detected rather than merely discouraged. Here is the wiring, the Firecracker boot args, and an honest account of what it does not cover.
Ajay KumarSigning and Attesting microVM Templates
A signature over only the rootfs leaves the snapshot unsigned — and the snapshot is a fully booted kernel's memory, restored on every create without ever executing a boot-time integrity check.
Ajay KumarSaying No Correctly: Admission Control for Sandbox Fleets
Almost no capacity bug is 'we ran out of RAM.' It is either a yes you could not fit, or a no you did not have to give — and the second one is more expensive, because it looks like your platform is broken.
Ajay KumarGolden Images vs Snapshot Baking
A golden image removes install time. A snapshot removes boot time. Those are different costs, which is why the answer is almost always both — and why the snapshot quietly freezes your RNG, your clock, and anything that was in RAM at bake time.
Ajay KumarRunning Firecracker Under Nomad
Nomad will happily place a Firecracker microVM on a host, restart it, and register it in Consul. It will not clone your rootfs, build your network, restore your snapshot, or notice that the guest inside your perfectly healthy VMM process has been wedged for ten minutes.
Ajay KumarThe Best Multi-Tenant Isolation Platforms in 2026
You are not shopping for a sandbox. You are shopping for an isolation boundary between one paying customer and the next — and the right one depends entirely on what your tenants are allowed to do.
Ajay KumarScanning Package Postinstall Scripts in a microVM
The only reliable way to find out what an install script does is to run it — which is the exact thing you were trying to avoid. So run it somewhere disposable, with the network turned off, and let the package tell on itself.
Ajay KumarRunning Customer Trading Strategies in Isolated microVMs
Your customers write strategies. Your infrastructure runs them, next to each other, holding credentials that place orders. This is the untrusted-code problem with a P&L attached.
Ajay KumarIsolating Customer-Managed Key Operations in microVMs
You shipped BYOK to close an enterprise deal. Now your process unwraps tenant A's data key on a heap that also serves tenant B — and a heap dump is a remarkably efficient way to violate a data-processing agreement.
Ajay KumarHow to add human approval to agent code execution
The gate that asks about everything gets clicked through by Thursday. The useful design asks about almost nothing, because isolation earns you the right not to ask.
Ajay KumarThe best AI agent guardrail tools in 2026, by layer
Most teams buy one guardrail product and think they're covered. The layers solve disjoint problems, and only one of them is deterministic.
Ajay KumarHow to run SWE-agent in a sandbox
SWE-agent hands a language model a shell over your repository. On a laptop, that shell also reaches your SSH keys. Here is the isolated version.
Ajay KumarHow to sandbox code from Cline and Continue
The approval prompts work — for about a day. Then you turn on auto-approve, and a model has an unattended shell next to your cloud credentials.
Ajay Kumar