Engineer at Salesforce (previously Informatica), based in Bengaluru. I debug large distributed data platforms in production and build the backend tools that make that faster. I like measuring things before optimizing them.
Things I've built
- inferj: GPT-2 inference in a single Java file, no dependencies, with its own tensor library and KV cache
- car: distributed log processing and analysis with pluggable stores and cluster processing
- sdfs: small distributed filesystem and job scheduler, run on 70 nodes
- metmon: metrics monitor for Hadoop components (HBase, HDFS, YARN) with a metrics sink, REST server, pluggable stores and a Grafana plugin
- tensor.benchmark: code, perf data and tools for my blog series "Taking JVM on a performance ride"
- active-trace-ids: servlet filter that turns on debug logging per request through flags in the trace id
- asm-vis: x86 assembly control-flow-graph visualizer that runs in the browser
Writing: JVM internals, performance and ML systems at mmpataki.github.io
Open source: a merged fix in Apache Hadoop (HADOOP-17081), and pull requests to Apache HBase (a Prometheus metrics endpoint), the MongoDB JDBC driver, kubernetes-client/java and llama.cpp

