1. Python web frameworks
http://www.techempower.com/benchmarks/#section=data-r9&hw=i7...
Carefully configured Flask can get up to 1/3rd the speed of other web frameworks. But misconfigure it, use an ORM, or if you're on django, you're serving 1/10th as many requests/sec with significantly higher latency than other languages.
2. Templating
https://stackoverflow.com/questions/1324238/what-is-the-fast...
Template rendering takes an appreciable amount of time. I had to replace django's (slow as shit) stock templates with jinja2 on a project because it was taking tens of seconds to render a very large page. Jinja is far faster, but it's still slow enough that I had to add a caching layer to a website that shouldn't need one.
3. Numerical shit
I implemented a video processing algorithm from a paper for a computer vision class. Python's scientific and numeric libraries were sick and made prototyping delightfully easy--then the final runs on the full dataset took forever. What I re-implemented in C++ was >10x faster. Vectorization and other tweaks can make python numerics go faster, but it's a pain in the ass. One of the goals of the Julia scientific programming language is to make loops fast enough that code doesn't have to be manually vectorized.
I prototype in python and use zeromq so i can simply insert a new piece of code in c++ when the time is right. Throughput generally increases 10x or more for the same algorithm
