[ronancremin]

Speaking for DeviceAtlas in particular I can assure you that detection reliability is better than most assume. The stated 99% detection accuracy rate isn’t marketing speak—it is continously verified on real traffic from all around the world.

We measure two main aspects of performance:

- Per-request detection overhead: Willy’s preferred way to measure impact on HAProxy performance is the per request overhead added. DeviceAtlas adds a few µs per request. Typically this does not cause much of an issue. Example: if a load balancer is serving 20,000 requests per second with 80% CPU, that's 40µs of CPU time per request on a machine that can go up to 50µs. If we add 4µs to that we reach 88% of CPU under the same load, and the end user performance is not degraded in any meaningful way.

- Memory footprint: DeviceAtlas lets you configure the per-device property set to tailor the memory and performance impact. The resulting memory impact ranges from about 12MB to 100MB.