Omar Faroque (@faroque_eee) header
Omar Faroque (@faroque_eee) profile picture

SWE at @Microsoft AI ex @meta, ex @aws, ex @apple Currently learning more on Agentic application and LLM Security.

San Francisco
Joined April 2015
Why CPU utilization based Load balancing increase latency and errors? In modern application system, compute runs on VM with many other VM on a host machine. We call other VM that runs with our application antagonist. Now come to an example:
1
7
1,564
One point after @elonmusk bought tweeter, many left the tweeter. But now many more people are coming and becoming active. What makes this return?
2
252
Everyone(listed) is talking about open weights model, but the listed companies here does not have a good model yet ironically. So they are just begging for free good model. Also, they are trying to create more enemy for USA or not, I don’t know. I think it will give China and other USA competitors an advantage.
Open-weight models are essential to a healthy AI ecosystem. Together with others across our industry, we are outlining a path for open-weight models to strengthen American competitiveness and expand economic opportunity, while protecting national security. microsoft.com/en-us/corporat…
1
229
I was discussing with claude about Multi level security particularly about label. Look at this example below, how claude makes an example of a codeword named "Iran-Ops" and why it did ? I am just curious.
1
172
I was using opus 4.8
107
How a simple #caching strategy if not careful can brought a black swan events or #Metastable failure for your service. I wrote a summary based on a paper(sigops.org/s/conferences/hot…). nanofaroque.github.io/person…
197
I often asked why we should learn in the age of AI. Here is my thought in this article, please read it and share your thoughts.
1
1
154
Do you remember when you joined X? I do! #MyXAnniversary Wow
102
Nice explanation by Claude: Lock = a key to a room (only one key, only the holder can leave and hand it back) Semaphore = a bouncer with N wristbands (anyone leaving can hand their wristband to the next person)
106
Just for the fun and discussion with my colleague and friends I(Claude make it nicer) wrote this #GAS #theorem for AI. nanofaroque.github.io/person…
90
AGI to ASI Artificial General Intelligence to Artificial Super Intelligence
82
I was looking into ggml-org/llama.cppPublic and a file: src/llama-model.cpp. I found the code snippet below and it is funny and means: "I don't know how many billions parameter you want to use"
67
Such prestigious conference paper can have typo too. @ACMSIGCOMM conferences.sigcomm.org/hotn…
202
Nice write up. However logs can be used for metrics dashboard and monitoring as well. As an example, count for some special type of error message. However that should be at the very early at the source or when the log stream reach to the log processor.
How to design an efficient logging system Most teams don’t have a logging problem. They have a retention and indexing problem. The fix is to treat logs like data: Hot logs for fast debugging, cold logs for cheap storage, and a lifecycle that automatically moves or deletes everything. 1. Split logs into Hot vs Cold. Hot logs are what on-call actually needs: the last 7–30 days, searchable, low-latency. Cold logs are everything older: kept for audits and rare investigations, optimized for cost. 2. Filter and structure at the source (Not in the log tool). Your app should emit structured logs (JSON) with consistent fields like service, env, level, request_id, user_id, status, latency_ms. Use an agent (Fluent Bit / Vector) to drop debug logs in prod, sample noisy endpoints, and redact PII early. The cheapest log line is the one you never ingest. 3. Use a buffer so spikes don’t degrade your pipeline. Ship logs from the agent to a stream/buffer (Kafka / Kinesis / PubSub). This absorbs burst traffic, protects downstream sinks, and gives you backpressure instead of dropped logs when something goes wrong. 4. Fan out to two sinks: Datadog for Hot, S3 for Cold. Send the curated, high-signal subset to Datadog for search, dashboards, and incident response. Send the full (or fuller) stream to S3 for long-term retention. Datadog is for debugging velocity. S3 is for durability and cost control. 5. Store cold logs in S3 in a query-friendly format. Write logs to S3 partitioned by date and service, like dt=YYYY-MM-DD/service=foo/env=prod/. Store as compressed Parquet when possible (or gzipped JSON early on). This makes later investigations cheap with Athena/Trino without re-indexing everything into an expensive log system. 6. Enforce retention with lifecycle rules, not discipline. Set S3 lifecycle policies: Keep hot searchable logs only 7–30 days in Datadog. In S3, transition older logs to Infrequent Access or Glacier, and expire them after your compliance window (90/180/365 days). Unlimited retention is just unbounded cost. 7. Make “Cold queries” a first-class workflow. When someone needs older logs, don’t pretend they should be instantly searchable in Datadog. Provide a simple path: Athena saved queries, a small internal tool that fetches partitions by request_id/time range, or an on-demand rehydration job for rare cases. 8. Don’t use logs as your alerting backbone. Alerts should come from metrics and traces (Golden signals). Logs are for context after an alert fires. This reduces indexing pressure and keeps logging costs stable as traffic grows. If you build it this way, observability stays great for engineers, but predictable for finance. You debug fast from Datadog, you retain safely in S3, and you stop paying for data you almost never query.
1
106
In ultra low latency and financial service like #stockexchange, I saw system that write sell and buy order into sell queue and buy queue. How to make sure reliability of the transaction incase of machine failure? #SystemDesign
1
80
I was thinking thinking to combine: WAL+Synchronous replication. That WAL writes into to some kind of database. Is this thinking in right direction or I am missing some key parts?
60