Kepiye sampeyan nangani hot key utawa hot row ing contention abot?

Question

IT Interview Practice · Accepted Answer

**Mandheg nyalurake kabeh lalu lintas menyang siji key: pisah hot key dadi N sub-key (sharded counter), gabungake pamaca duplikat, ngladeni saka cache lokal, lan batch tulisan supaya siji row ora diupdate saben request.** Siji counter "likes" ing postingan viral bisa nampa 50rb tulisan/detik — siji row dadi bottleneck serialisasi amarga saben transaksi rebutan lock sing padha.

## Fan-out tinimbang siji row sing diperebutake

```text
                 50k inc/sec on post:123
                          │
             hash(request) % N   (split the key)
        ┌────────┬────────┬────────┬────────┐
        ▼        ▼        ▼        ▼        ▼
   likes:123:0  :123:1  :123:2  :123:3  :123:4   ← sharded counters
        └────────┴───┬────┴────────┴────────┘
                     ▼
             SUM on read = total     (read fans in)
```

## Teknik

- **Sharded counter / pemisahan key.** Tulis menyang `likes:123:{0..N}` sing dipilih kanthi acak; cacahan sejati yaiku `SUM`. Iki ngowahi siji hot row dadi N row sing luwih adhem, ngurangi contention lock kira-kira N kali. Pamaca mbayar biaya fan-in cilik.
- **Request coalescing (single-flight).** Yen 10rb request nyuwun key sing padha bebarengan, ben **siji** ndemek DB lan liyane ngenteni asil sing lagi mlaku kuwi. Nyegah stampede cache-miss ngremuk row.
- **Cache lokal (in-process).** Cache nilai panas ing saben node app kanggo TTL cendhak (contone 1 detik). Kanggo key sing panas maca, kuwi ngempalake ewonan pamaca/detik dadi siji refresh — nampa rada basi.
- **Batching tulisan.** Buffer increment ing memori lan flush `+= 500` sepisan saben detik tinimbang 500 update siji-siji. Siji update row saben detik tinimbang 500 ngilangi contention babar pisan.

## Sudut pandang wawancara

Pewawancara pengin ndeleng sampeyan mendiagnosa **ing ngendi contentione** lan cocogake solusi karo read-hot vs write-hot. Read-hot → cache + coalescing; write-hot → sharded counter + batching. Jawaban apik ndhaptar teknik; jawaban apik banget nyathet yen sharding nge-trade **biaya maca kanggo skalabilitas tulisan**, lan yen batching/caching nge-trade **kesegeran kanggo throughput**.

**Trade-off & jebakan:** sharded counter ndadekake total real-time sing pas luwih angel (sampeyan njumlah shard) lan njlimet logika "kurangi ngisor nol". Cache lokal nyimpang antar-node — apik kanggo cacahan like, salah kanggo saldo bank. Batching ngambil risiko ilang tulisan sing di-buffer yen node crash sadurunge flush, mula cocog kanggo counter idempotent/perkiraan, dudu dhuwit. Pilih mitigasi sing diidini kabutuhan konsistensi data.

Fan-out tinimbang siji row sing diperebutake

Teknik

Sudut pandang wawancara

Pitakon sing gegandhengan