Tell us your success stories using memray!

[pablogsal]

If you have used memray and found it helpful, please leave a comment below with your success story. We would love to hear from you and know that our efforts are making a difference. Maybe you used memray to detect a memory leak in a C extension or to detect how memory was being used in your large application. If so, you can tell us how did memray help!

Knowing how we managed to help your use case will help us track trends and target areas for improvement, prioritize new features and development and identify potential bugs or areas of confusion.

Thank you for taking the time to provide feedback! ❤️

[JelleZijlstra]

I maintain a large Python application at Quora. It runs a couple of pre-warming tasks at startup to improve performance at runtime. I used memray to figure out which parts of the warmup sequence were using memory. Fixing a few of the bottlenecks helped us achieve a significant cost reduction.

[tonybaloney]

I work on a Python C extension that is designed for performance. I successfully used memray to identify memory leaks (refcount leaks) of Python objects created in the extension module whilst developing it. We're using memray for Quality Assurance testing on this library that's being used in a hot-loop executed billions of times a day.
All the issues identified were done during the development cycle instead of identifying memory usage in production environments.

[pamelafox]

Pics for proof! (From that project)

[jmg-duarte]

I work on a data pipeline system (akin to Prefect, Dagster, Airflow, etc) and out of the blue, the system worker started crashing due to OOM errors. While we were able to identify that it was a memory leak on the LZ4 part of the system, we couldn't track down whether it was on the python-lz4 C bindings or the actual LZ4 library.
With the collaboration of the LZ4's author, I isolated the cause to python-lz4 C bindings (lz4/lz4#1096).
Afterwards, using memray (with the native flag) I was able to directly track the line that allocated the leak memory - memory wasn't being freed in the error path - adding a PyMem_Free resolved the issue (python-lz4/python-lz4#255).

[bgyori]

I lead the development of several open-source scientific software projects for computational systems biology. We recently used memray to profile the memory usage of our Gilda software and successfully identified where it can be streamlined. I found the live monitoring as well as the HTML-based visualization features of memray particularly useful.

[hmcezar]

I work on the development of HyMD, a scientific software for molecular dynamics simulations using a hybrid particle-field approach. Using memray we were able to track a huge memory leak from OpenMPI that leaked into our application, causing many simulations to crash after a while. memray was an indispensable tool to help us fix this problem.

[pquentin]

I co-maintain urllib3, the package with the most downloads on PyPI. Anything we do has a big impact, for better or for worse. As part of urllib3 2.0, we wanted to fix how read() works with compressed data to be more intuitive. This required adding a buffer, but our first attempt doubled the memory usage. Armed with memray and its intuitive flamegraphs, I managed to come up with a solution that used the same memory as before. Without that reduction we could not have shipped this improvement. And with pytest-memray we're now confident that there won't be any regressions!

[farhanhubble]

I was handed over a massive natural language processing pipeline that was consuming a whopping 250GB of RAM!! I looked at the models. Quite a few of them but the largest one was about 1GB. I tried memray and looked at the summary. The peak memory utilization was just a few GBs. This led me to believe that something was allocating and deallocating memory too fast and the Python interpreter was not able to return the memory back to the system in time. Eventually found and fixed the code that was repeatedly loading a machine learning model on every call.

Memray is fast, beautiful and accurate. I wish I could do a t-shirt with a snapshot of memray --live to thank the devs.

[maartenbreddels]

I've described at widgetti/solara#172 how I used memray to find the area where a memory leak was in Solara and/or Reacton, found the memory leak (global reference) using objgraph and then confirm the fix using memray again.

[ankush]

Memray played a huge role in our effort to reduce memory footprint at @frappe. I've used many different memory profilers in past but none come close in terms of usability and simplicity. We were able to slash anywhere from 10-35% memory usage (depending on the configs) and that translates to gigabytes of savings across servers that serve 6k+ sites 😄

--follow-fork was quite useful feature when dealing with worker processes that use forking.

references:
frappe/frappe#21467
frappe/frappe#21473
frappe/frappe#21474
frappe/frappe#21475

[tvalentyn]

Memray was instrumental in pinpointing a source of a tricky memory leak in Apache Beam that I was investigating: apache/beam#28246. I tried several tools and Memray is by far the fastest and the most usable tool I tried. I haven't used memray before and originally tried memray summary command which wasn't informative, and made me dismiss Memray. I decided to give Memray another try a couple of weeks later after very frustrating experiences with many other, mostly older tools, and discovered that memray table command does wonders for hunting down my leak. I later used Memray to simplify the repro and Memray dramatically reduced my feedback loop during debugging.

[terapyon]

I wrote an article about Memray in Japanese last year. Title is "メモリプロファイラ「Memray」の解説", it is "Explanation of the memory profiler “Memray". https://gihyo.jp/article/2023/06/monthly-python-2306
This is one of famous online media in Japan.

Then, I gave a talk it at PyCon APAC 2023 in Japanese.
https://2023-apac.pycon.jp/timetable?id=FHRP3H

[tojimahammatov]

I work on developing services which some of them perform AI model inference. We noticed a memory leak issue in one service and detected 2 reasons for it. Although we fixed one, tracing the other issue was quite difficult. Memray helped us to identify the bug accurately by providing nice reports.