Comment by supriyo-biswas
19 hours ago
A major player in this space is apparently looking for people experienced in scraping without using browser automation. My guess is that not running a browser results in using far fewer resources, thus reducing their costs heavily.
Running a headless browser also means that any differences in the headless environment vs. a "headed" one can be discovered, as well as any of your Javascript executing within the page, which significantly makes it difficult to scale your operation.
My experience is that headless browsers use about 100x more RAM, and at least 10x more bandwidth and 10x more processing power, and page loads take about 10x as long time to finish (vs curl). Though these numbers may be a bit low, there are instances you need to add another zero to one or more of them.
There's also considerably more jank with headless browsers, since you typically want to re-use instances to avoid incurring the cost of spawning a new browser for each retrieval.
Is it possible to pause a VM just after the browser has started up? Then map it as copy-on-write memory and spin up many VMs from that "image".
Your comment is interesting and there are some people doing work on this although not specific to browser automation, e.g. AWS Lambda SnapStart is just them trying to boot your Java Lambda code and freeze the Firecracker MicroVM's snapshot and then starting other Lambda functions from there.
However, even with a VM approach, you tend to lose out on the fact that you can make 100s or 1000s of requests on a small box (~512 MB) every second if it's just restricted to HTTP(s). However, once you're booting up a headless browser, you're probably restricted to no more than loading 3-4 pages per second.
... but then you have even larger overhead, as well as the added layer of complexity from managing VMs on top of headless browsers.
On the other hand you need to be able to do basics like match the headers, sometimes request irrelevant resources, handle malformed documents, catch changing form parameters, and other gotchas. Many would just copy the request from the browser console.
The change rate for Chromium is also so high that it's hard to spot the addition of code targeting whatever you are doing on the client side.