When optimizing, treat performance as a design constraint: prevent contention/cache inefficiency, ensure server-side work is bounded per call, and keep tests from becoming slow bottlenecks. 1) Cache-line & contention hygiene (C structs) - Align/pad frequently-updated counters to avoid false sharing. - Avoid mixing rarely-used “cold” fields into the same cache footprint as hot ones; move them deeper in the struct. - Prefer per-thread stats to share less with other threads; if needed, split “hot counters” into their own cache-line-aligned struct. Example pattern: ```c typedef struct __attribute__((aligned(CACHE_LINE_SIZE))) { redisAtomic long long io_reads_processed; redisAtomic long long io_writes_processed; /* other hot, per-IO-thread counters */ } IOThreadStats; typedef struct { /* hot per-IO-thread fields */ IOThreadStats stats; /* each thread updates its own cache line */ /* cold / infrequently read fields after */ int unused_cold_field; } IOThread; ``` 2) Bound work for collection-style commands - For commands that iterate over IDs supplied by the client, ensure per-ID processing is O(1) with a small constant, and rely on the existing batching convention (client chooses how many IDs to send). - Only introduce server-side COUNT/batching knobs if they meaningfully prevent a real bottleneck or reduce complexity (not just to mimic client-side batching). 3) Keep tests fast and deterministic - Avoid long fixed sleeps for timing/cron behavior. - Use polling/`wait_for_condition`-style loops and reduce durations/timeouts so CI runtime stays low.