Build1 publisher2 min readPublished
Cloudflare opens its internal Quicksilver store to customers as Workers KV Instant
Cloudflare's Workers KV Instant puts the KV API on Quicksilver, claiming p99 reads 100 times faster and writes global within about 250ms. At $100 per stored megabyte a month, it is priced for small sets of flags and settings that rarely change.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Cloudflare had never offered Quicksilver to customers before, though it says they asked for years; the new mode runs on Quicksilver v2.
- In absolute terms, Cloudflare reports KV Instant reads resolving in under two milliseconds at p99, with p95 latency measured in microseconds.
- Cloudflare prices KV Instant reads 60% below classic Workers KV, while storage and Class A operations cost significantly more.
- Metadata is not supported in instant mode: getWithMetadata calls always return null, and writes cannot attach metadata.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- cost Every megabyte kept in an Instant namespace costs $1,200 a year, a price that confines the mode to small configuration sets and sends anything that grows with users back to classic KV.
- exposure Code ported from classic mode that fetches values through getWithMetadata will get null on every call, and a handler that treats null as a missing key will serve its fallback without raising an error.
- decision Teams have to sort keys by size and write rate before picking a mode, since Cloudflare's own guidance sends large or frequently updated data to classic Workers KV.
The design difference is where a value sits when a Worker asks for it. Cloudflare says KV Instant propagates data to every one of its locations [15], and its replication notes count more than 300 edge locations at publication [7]. Reads come from a local copy, so there is no cold read penalty and no TTL to wait out [1]. Quicksilver has done this inside Cloudflare since 2020. The company wrote that "nearly every request to Cloudflare looks up at least one key in Quicksilver" [2].
Full replication explains the price sheet. Every stored byte is held at every location [1], so reads are the cheap resource and bytes are the expensive one. Storage is billed at $100 per MB per month [18]. Taking a gigabyte as 1,000 MB, one gigabyte would cost $100,000 a month [3]. Keys are capped at 300 bytes [19].
The 100x is a ratio against classic mode [4]. For it to transfer to a given app, that app's classic p99 has to be made mostly of the cold reads Instant removes. I'd expect the gap to be widest for keys read rarely at any one location and narrowest for a key every request already hits.
The write ratio rests on a coarse baseline. Cloudflare says writes reach the edge over 20 times faster [6], but a table note says it does not have sub-second fidelity for replication lag in classic mode [8]. The 20x compares a quarter-second figure against a classic number Cloudflare cannot resolve below one second [6][8].
"Immediate" in this post means about 250ms for 99% of writes [6]. The post does not give a figure for the other 1%. Cloudflare's own example fits inside that window easily: a homepage that switches at 10:13 AM when a product is announced on stage, gated by a flag read on every request through `env.APP_CONFIG.get("new-product-launched")` [10].
Moving a flag namespace over is cheap, and keeping the existing API was the right call. KV Instant keeps get(), put(), list() and delete() [11]. The mode is chosen when the namespace is created, by passing "mode": "instant" [12]. A list call returns every matching key in one response, with no pagination [14], so cursor handling written for classic mode has nothing to do.
In my context, a few kilobytes of flags and kill switches read on every request, this is the right tradeoff. Cloudflare scoped it the same way. "It's not for every type of data," the company wrote, naming infrequently updated application configuration, the same data it keeps in Quicksilver itself [9].
What to watch
- A published figure for the slowest 1% of KV Instant writes, beyond the roughly 250ms Cloudflare gives for 99%.
- Whether Cloudflare documents converting an existing classic namespace, since instant mode is chosen at creation.
- Independent p99 measurements comparing classic and Instant reads on customer workloads, to test whether the 100x ratio holds.