Cloudflare Opens Quicksilver to Customers as Workers KV Instant, Inside a 1 MB Envelope

Cloudflare recently made Quicksilver available to customers for the first time, as a new mode for Workers KV called KV Instant. Reads resolve in 1.62 ms at the 99th percentile, compared with 287 ms for classic KV. What determines whether a team can use it isn't the latency, but the envelope around it.

Quicksilver is Cloudflare’s own configuration store, introduced in 2020. Nearly every request to Cloudflare looks up at least one key in it, Rob Sutter writes, and it has never been a customer-facing product. Brendan Irvine-Broque, who works on Cloudflare’s developer platform, announced the change on X:

we have this crazy key-value store at Cloudflare called Quicksilver

instantly updates keys across our whole network. microsecond reads from every location. stupid fast

our customers have asked us for years to make it available

we finally did – in a way we think is quite simple - as a new mode for Workers KV

The published figures, measured to over 300 edge locations:

Mode	p99 reads (all)	median write replication	p99 write replication
Instant	1.62 ms	107 ms	256 ms
Classic	287 ms	under 1 s	4.38 s

Cloudflare notes it has no sub-second fidelity for replication lag in classic mode, so the classic median and p95 figures are upper bounds rather than measurements. p95 reads in Instant mode are reported in microseconds.

The API is the Workers KV one, with three differences. Mode is set at namespace creation with a

"mode": "instant"

attribute, and cannot be changed afterward. Metadata is unsupported, so

getWithMetadata

always returns null. And

list

returns every matching key with no pagination.

Then the limits. A namespace holds at most 10,000 key-value pairs and one megabyte in total, with keys up to 300 bytes. Writes are capped at one per namespace per second, which Cloudflare contrasts with classic KV’s one write per key per second, on the grounds that a single system of record "makes it easier for you to reason about update order in your application." That difference is the one most likely to catch a team migrating: a pattern classic KV permits across many keys concurrently becomes a single serialised queue.

Pricing points the same direction, and Cloudflare says so plainly: "storage and Class A operations are significantly more expensive than Workers KV."

	Reads (Class B)	Writes, deletes, lists (Class A)	Storage
KV Instant	$0.20 per million	$0.10 per operation	$100 per MB, per month
Classic KV	$0.50 per million	$5.00 per million	$0.50 per GB, per month

Reads cost 60% less. A write costs ten cents, against five dollars per million in classic mode. Storage runs $100 per megabyte per month, and since a namespace tops out at one megabyte, that is also the ceiling. Multi-key reads bill per key, so requesting three keys in one

get

costs the same three Class B operations as three separate calls.

Irvine-Broque set out the reasoning in the same thread:

just like how we use Quicksilver ourselves, it’s for small data that is read often and updated infrequently

pricing reflects this - reads actually cost 60% less than even Workers KV

Writes cost what they cost, he added, because writing a key replicates it to hundreds of locations. This is a store priced so that you read from it constantly and write to it rarely.

Benjie Malinao, also on X, read the fit the same way:

This is the config store agents actually need. Sub-2ms p99 reads and global writes landing in about a quarter second at p99, on the same KV API. Feature flags and agent settings without waiting on a TTL. I’d move hot-path config over first.

Feature flags, routing tables and agent settings fit that shape. Session state, user records and anything with a write path do not.

The comparison practitioners reached for was Vercel. Developer Shawn Mclean asked Irvine-Broque directly:

I’m looking into using Vercel Global Config for a middleware read. They claim 15ms. If I move my workloads to CF, are you saying I can read from middleware for a platform tenant domain mapping in 2ms?

The answer did not quite confirm it:

even p50 of Workers KV (non-instant) from 2 years ago was ~5ms. KV Instant is even faster than that (see numbers in the post)

Cloudflare’s own worked example is a launch flag read on every request and flipped at a keynote, which is precisely the workload: one key, one write, read everywhere immediately.

Classic Workers KV, which launched during Birthday Week in 2018, remains the recommendation for large or frequently updated data. The two modes now sit under one API with opposite economics.

KV Instant is in private beta with a sign-up form, and Cloudflare says it plans to open it up more widely. No general availability date is given.