Tuning for HGetAll

[ahapeter]

I have a hash key with 128 field; each field size arround 140 bytes
It's take about 50ms (p90) to fetch from redis. It's even slower than direct query from postgres - which just take around 1ms.
I also tested with go-redis - take arround 10ms
I think it's about tuning.

[rueian]

Hi @ahapeter,

1ms sounds impressive but it could become slower when your table grows.

The 50ms could be caused by auto pipelining's head-of-line blocking. To improve latency over throughput, you can try DisableAutoPipelining=true globally or isolate latency-sensitive operations in a dedicated connection (client.Dedicated()).

But to compete with the 1ms from postgres, I would suggest using client-side caching client.DoCache() if all values in your hash key are seldom changed.

[ahapeter]

hi @rueian
I was try to set DisableAutoPipelining=true, increase ring buffer factor, per connection cache size, read/write buffer
It's not change much.

also, i am confused, maybe not related to rueidis, write to hashset (to keys, fields) also slower than postgres.
In my machine: postgres, redis and go app run on same machine. AMD 8700G, 16G RAM, Ubuntu 22.04

• redis (via rueidis): ~20k rps
• postgres: ~75k rps
  To be fair; when write to postgres have 8 worker that do the batch write. I was batching write to redis before (when use go-redis) but when switch to rueidis, I just leave it run with default configs (auto pipeline, flush interval 100us)

any ideas from you?

[rueian]

I was try to set DisableAutoPipelining=true, increase ring buffer factor, per connection cache size, read/write buffer
It's not change much.

When DisableAutoPipelining=true, the ring buffer size does not matter because it is only used by auto pipelining. But BlockingPoolSize matters since it now uses the connection pool to serve concurrent requests. If you keep constant 8 workers, then the default BlockingPoolSize is fine.

per connection cache size only matters if you use client.DoCache().

auto pipeline, flush interval 100us trades latencies for more throughput.

Both the default read/write buffer sizes are 0.5MiB which is enough to hold your large HSET.

write to hashset (to keys, fields) also slower than postgres.

~20k rps throughput is expected for writing your large HSET. I also got the same number on my laptop. My Redis already ran at full speed. Remember that Redis is a single-thread server. You may try dragonflydb or keydb for utilizing multi-core for higher throughput.

However, 50ms latency is quite unexpected on a local machine. On my laptop, I got 0.3ms for p90.

[ahapeter]

thank you so much