Uber Cuts Translation Cache Memory by 90%

Uber's engineers redesigned a Go-based translation cache, reducing its initial memory footprint by more than 90 percent while keeping cache hit rates above 99.9 percent.

Updated on Nov. 20, 2026 in Software

Uber Cuts Translation Cache Memory by 90%

Uber redesigned its Go-based translation cache to cut its initial memory footprint by more than 90 percent. The hybrid system keeps translation keys and frequently accessed values in memory and stores the rest on disk, serving millions of translations per second.

Why it matters

Garbage collection consumed more than 20 percent of CPU in some services, with translations accounting for roughly 70 percent of allocated objects, and cache refreshes had caused some services to run out of memory. The redesign targets those performance bottlenecks in high-throughput backend systems.

The implementation uses 256 shards, each with its own read-write mutex, in-memory entries and disk file, with p99 latency for disk-backed cache misses staying below 100 microseconds. Translation datasets can contain two to three million keys, and the figures reported come from Uber's own internal measurements.

The players

Uber

Uber is a global ride-hailing and delivery technology company whose engineering teams build large-scale backend infrastructure.

The details

Uber's cache uses a time-based global least recently used policy for in-memory entries, and during refreshes it preloads translations accessed in the previous 15 minutes. Previously, both the old and replacement caches occupied memory at once during refreshes, which contributed to out-of-memory failures in some services.

Timeline

  1. Article on the Uber translation cache redesign was published November 20, 2026.

The Tech Race

Uber's replacement of Go's sync.Map with a map protected by an RWMutex follows a common pattern of large operators tuning standard library primitives for their own throughput demands. The change shows how memory and CPU trade-offs in concurrent Go code are being revisited at hyperscale.

For users of Uber's apps, the change is invisible, but it indicates how backend efficiency can reduce infrastructure costs. Engineers running large Go services can apply the same measurement-driven approach to memory and garbage collection tuning.

The takeaway

Profiling garbage collection and memory allocation under production load can reveal bottlenecks that synthetic tests miss. Choosing data structures that match the access pattern, such as sharded maps over generic concurrent maps, can yield large efficiency gains.

Further reading

Explore more in Software for related coverage of developer tools and backend systems.

Source note: This article includes information reported by InfoQ.