Fast, cheap, robust: you don't have to pick two

Abbas Ali ·

There is an old rule in engineering, usually drawn as a triangle: fast, cheap, good — pick two. You can have it fast and good, but it won't be cheap. Fast and cheap, but not good. Good and cheap, but not fast.

For an in-memory database, "good" means robust: it keeps your data and keeps answering when a machine dies. And the managed market has followed the rule closely. The fastest options cost the most. The cheap ones run on a single node, or ask you to accept that a failover may lose writes. The robust ones charge for every replica twice over.

We built KiviDB Cloud on the bet that the triangle is a habit, not a law. Here is what each corner looks like, with the numbers behind it.

Fast, cheap and robust drawn as the corners of a triangle, all three ticked, with KiviDB at the centre.

Fast

KiviDB is a Redis-compatible engine built to use every core on the machine. On the same hardware and the same memtier_benchmark parameters as Redis 8.6 and Dragonfly v1.37.0, KiviDB v1.0.5 handled:

  • 18.5 million pipelined writes a second — about 31× Redis and 3.3× Dragonfly — with a p99 of 0.35 ms.
  • 25.9 million pipelined reads a second — about 24× Redis and 2.9× Dragonfly — with a p99 of 0.32 ms.
  • 27.9 million mixed reads and writes a second, with a p99 of 0.34 ms.

Bar charts of pipelined SET, GET and mixed throughput: KiviDB at 18.5M, 25.9M and 27.9M ops/s with p99 latency of about a third of a millisecond, against Dragonfly and Redis.

The p99 is the number to watch. Throughput tells you how much a machine can do; tail latency tells you how your slowest requests feel. Redis's p99 on the pipelined write test was 20 ms. KiviDB's was a third of a millisecond.

These are pipelined, synthetic numbers, measured on AWS Graviton3 (c7gn.12xlarge for each store, with the load generator on a c7gn.16xlarge in the same zone). Your workload will differ, so the full commands are in our benchmark guide — run them yourself.

Cheap

Speed is only half of the cost story. The other half is how much machine you need to hold your data safely, and how a provider charges you for it.

We compared the list price of a highly available database — one primary and one replica — across the managed options, at 5, 25 and 100 GB:

Monthly price per GB with one replica at 5, 25 and 100 GB: KiviDB Cloud lowest at each size, against Dragonfly Cloud, Amazon ElastiCache, Memorystore and Redis Cloud.

Dataset, with a replicaKiviDB CloudCheapest alternativeMost expensive alternative
5 GB$95.54/monthDragonfly Cloud, $143.00Redis Cloud Essentials, $258.00
25 GB$392.70/monthAmazon ElastiCache, $480.05Redis Cloud Pro, $1,708.00
100 GB$1,570.42/monthAmazon ElastiCache, $1,917.86Redis Cloud Pro, $3,416.00

That is 18–77% less than the alternatives, depending on size and provider. For each competitor we took its cheapest published configuration that holds the dataset with a replica, on-demand list prices in US regions, read on October 5, 2026. KiviDB is priced as Pro, standard throughput, on AWS — and the number is the same one we bill from.

Where does the difference come from? Not from cutting corners:

  • Memory-optimised machines, sized with headroom. Your data fills at most 80% of the machine's memory. The rest is room the engine needs — for the operating system, for background saves, for replication buffers — so you are not one backup away from running out.
  • An engine that does more per core, so a database needs a smaller, cheaper machine to reach the same throughput.
  • Billed by the hour. No reserved instances, no annual commitment, no minimum term.

Robust

A database that is fast and cheap but loses your writes is not a database you can use. On KiviDB Cloud Pro, robustness is the default, not an add-on:

A primary in zone A replicating to a replica in zone B, backups alongside, and the failover sequence: primary lost, replica promoted at the same address, new replica built.

  • Replicas in other availability zones. Every write on the primary is replicated to a standby, and replicas are placed in a different availability zone from the primary wherever the network spans more than one. The loss of a machine — or of a zone — leaves a copy of your data ready to serve.
  • Automatic failover, same address. When the primary is lost, the replica is promoted and your endpoint keeps pointing at whichever node is serving. Your application reconnects to the name it already has. A new replica is then built in the background.
  • Persistence on by default. Writes go to disk as well as memory, so a restart is not a reset.
  • Backups you control. Export a .kdb file at any time, take disk snapshots on a schedule, and restore either in place or into a brand-new database.

We test this the hard way: we terminate a live primary that is holding data and watch what happens. In our failover tests the replica was promoted and serving writes within half a minute, with every key intact, and a fresh replica was rebuilt behind it without anyone touching anything.

On top of that: TLS for data in transit, private connectivity so traffic never touches the public internet, and the option to run your databases in your own cloud account if your data has to stay there.

Pick three

The triangle exists because, for most systems, every gain on one corner is paid for on another. A faster engine changes the arithmetic: when the database does more with each core, you can afford the replica, the headroom and the backups, and still come in under the price of the slower option.

That's the whole idea behind KiviDB Cloud. You shouldn't have to choose between a database that is fast, one that is affordable, and one that is still there when a machine fails.

Price your database — the configurator shows the monthly and hourly price for exactly the size, throughput and replicas you need.