samber-hot

Start with W-TinyLFU as the general-purpose default for mixed workloads. Switch only when measured hit rate misses your SLO: use LRU for recency-dominated patterns, LFU for stable frequency patterns, S3FIFO for large high-throughput caches, and ARC for self-tuning when patterns are unknown.

Chain WithLoaders on the builder with a function that batch-fetches missing keys, such as a database query. Concurrent Get calls for the same missing key share one loader invocation via singleflight deduplication, and you should check both err and found on every Get.

Estimate per-entry cost as struct size plus heap-allocated fields, key size, and roughly 100 bytes of bookkeeping overhead. Divide your memory budget by that item size and round down; measure item size with a test using runtime.ReadMemStats rather than guessing.

SetMissing panics at runtime if the missing cache was not configured in the builder. Add WithMissingCache(algorithm, capacity) for a dedicated negative cache or WithMissingSharedCache() to store missing keys in the main cache before calling SetMissing.

No, WithoutLocking and WithJanitor are mutually exclusive. Background expiration cleanup requires locking, so keep locking enabled whenever the janitor runs; WithoutLocking is only safe for single-goroutine access.

Add WithPrometheusMetrics(cacheName) to the builder and register the cache with prometheus.MustRegister, since HotCache implements prometheus.Collector. Track hit ratio with rate(hot_cache_hit_count) / rate(hot_cache_get_count) and treat values below 80% as a sign of undersized capacity or the wrong algorithm.