@yashig406

Devops Engineer | Former SDE Intern @microsoft | Learning & Growing 🙂‍↕️

 Delhi, India
yashigupta4623.github.io/Dossier/
Joined February 2022
  • Tweets2K
  • Following110
  • Followers86
697Media
starting again either to rise or to rise again!
2
249
When production breaks, your hands are busy fixing it - not writing the postmortem. So, I built 'Postmortem Pilot' check it out at : yashigupta4623.github.io/pos… @WeMakeDevs x @WisprFlow @kunalstwt @tankots @SahajGarg6 @superaayush
19
hoping!
18
The higher you climb the ladder, the less appreciation you get. Everything you do is kinda expected.
16
44
4
1,781
31,901
wrapping up #P99CONF 2026 by @ScyllaDB with @wayneariola #ScyllaDB #P99CONF @ScyllaDB @P99CONF
4
114
Takeaway from the #P99CONF 2026 keynote by @ScyllaDB: “The Vertical Scaling Wall, and How Valkey Gets Over It” by @reconditerose: Latency is a cliff, not a slope. In her single-threaded Valkey benchmark, p99 stayed under ~2ms up to ~140K req/s.
1
2
101
more replies
they had provisioned. - Vertical scaling (I/O threads) - Horizontal scaling (clustering) The flame graph showed most of Valkey’s time going to syscalls (epoll, reading from TCP buffers, writing responses), not to executing commands.
1
9
The old I/O threading worked in phases, so while commands executed, 5–6 extra cores sat idle. The fix: keep the I/O threads busy all the time. #ScyllaDB #P99CONF @ScyllaDB @P99CONF
17
Takeaway from “Can Everything be an LSM?” by @almoggavra at #P99CONF 2026 by @ScyllaDB: Every database comes down to 3 trade-offs: - Read amplification - Write amplification - Space amplification
1
5
114
more replies
The catch: ❌ Objects are immutable, so you can’t update one block .You pay per GET/PUT, so lots of tiny files gets expensive. Latency: ~150µs for an NVMe cache miss vs 20–30ms on S3 Standard That’s close to 100x slower.
1
2
51
An LSM tree isn’t one point on the read/write curve. It’s the knob that moves you along it. #ScyllaDB #P99CONF @ScyllaDB @P99CONF
4
18
Takeaway from "Memory is Slow, Disk is Fast" at #P99CONF 2026 by @ScyllaDB: Your file can be fully in the page cache and still be slow through mmap(). Cached in memory ≠ mapped in your process's page tables. Every miss means the kernel walks page tables in DRAM.
1
4
73
"It's in memory" doesn't mean "it's fast." SIMD loop over mmap'd data: 5.5 GB/s Same loop, data preloaded to RAM: 13 GB/s, the single-core max The bottleneck was never the loop. It was how the data got into memory. #ScyllaDB #P99CONF @ScyllaDB @P99CONF
2
26
Takeaway from the #P99CONF 2026 keynote by @ScyllaDB, "The Fastest Object Storage Client in the World" by @georg_kreuzmayr : 1. An S3 GET isn't one call. It's DNS + TLS + HTTP + TCP + async I/O stacked together. ...
1
2
62
2 . If you want a predictable p99, you have to own the stack, not just the SDK on top of it. What actually happens when you call S3? S3 / Azure / GCP ↓ DNS · TLS · HTTP ↓ TCP/IP · Async I/O Five layers, five chances for tail latency. #ScyllaDB #P99CONF @ScyllaDB @P99CONF
2
32
Takeaway from "Capturing Lightning in a Bottle" by @vjabrayilov7 at #P99CONF 2026 by @ScyllaDB: HotSpot's tiered compilation isn't free. C1 Tier 3 code gathers telemetry so C2 can optimize later. That telemetry lives in one MethodData per method, shared by all threads.
1
3
28
Under high concurrency, the profiling itself can become your bottleneck. #ScyllaDB #P99CONF @ScyllaDB @P99CONF
2
17
Takeaway from "Capturing Lightning in a Bottle" by @forked_franz at #P99CONF 2026 by @ScyllaDB: HotSpot's tiered compilation isn't free. C1 Tier 3 code gathers telemetry so C2 can optimize later. That telemetry lives in one MethodData per method, shared by all threads.
1
2
1
11
1,320
Under high concurrency, the profiling itself can become your bottleneck. #ScyllaDB #P99CONF @ScyllaDB @P99CONF
5
59
Load more