Need serious help scaling crawlee

I have an ECS instance with 4vCPU & 16gb RAM. My scaling options are the following:

Plain Text

    maxConcurrency: 200,
    maxRequestsPerCrawl: 500,
    maxRequestRetries: 2,
    requestHandlerTimeoutSecs: 185,

I am starting 4 of these crawlers at a time.
Here is a snapshot log:

Plain Text

{"time":"2024-04-15T00:09:08.818Z","level":"INFO","msg":"PuppeteerCrawler:AutoscaledPool: state","currentConcurrency":1,"desiredConcurrency":1,"systemStatus":{"isSystemIdle":false,"memInfo":{"isOverloaded":false,"limitRatio":0.2,"actualRatio":0},"eventLoopInfo":{"isOverloaded":false,"limitRatio":0.6,"actualRatio":0.106},"cpuInfo":{"isOverloaded":true,"limitRatio":0.4,"actualRatio":1},"clientInfo":{"isOverloaded":false,"limitRatio":0.3,"actualRatio":0}}}

Can anyone help me identify the correct settings so it is not maxed out?

29 comments

SSaurav Jain

Our team will reply you soon!

PPepa J

Hi
Did you set the right env variable for memory allocation ? https://crawlee.dev/docs/guides/configuration#crawlee_memory_mbytes

bbmax

do you think that is what's missing here? I do set CRAWLEE_AVAILABLE_MEMORY_RATIO=.80

bbmax

but how come currency is 1 when I have it set to max of 200?

PPepa J

It could be many things, like requests being enqueued one by one, I cannot really tell without seeing the code. Does the code work for you as expected when running locally?

PPepa J

by starting 4 of these crawlers, you mean like 4 separate processes? And it can take 0.8 * 16GB = 12.8GB => 4 * 12.8GB = 51.2GB total memory?

bbmax

No this is one instance of node on one ec2 instance. I start 4 of them asynchronously I.e .run() with separate queues.

bbmax

Thanks for diving deep with me. Happy to answer any other questions. It feels kind of like trial and error to me but I guess I really don’t understand why concurrency isn’t hitting at least double digits

PPepa J

So what happens if you try to run it with env variable set to CRAWLEE_MEMORY_MBYTES=13107 which is about 80% of 16GB?

bbmax

Memory doesn't seem to be the biggest issue here, CPU is the one maxing out, but, I guess I don't understand the PuppeteerCrawler:AutoscaledPool: state log well enough to know what the bottle neck is.

bbmax

but happy to try that if you think that is the problem.

bbmax

I figured it was CPU becuase: ,"cpuInfo":{"isOverloaded":true,"limitRatio":0.4,"actualRatio":1},

PPepa J

When you run it locally everything works properly?

bbmax

will test in a few.

bbmax

i also have a huge cpu/mem

PPepa J

I would like to distinguish if it is resources, website or implementation related issue. 🙂

bbmax

running on local which basically has unlimited cpu/mem.

bbmax

it seems pretty fast

bbmax

but the log that it returns is still single digits for conurrency,

bbmax

Plain Text

","currentConcurrency":7,"desiredConcurrency":8,

bbmax

here's an interesting thing from the ec2 instance:

Plain Text

{"time":"2024-04-15T01:36:05.694Z","level":"INFO","msg":"PuppeteerCrawler: Final request statistics:","scraper":"web","url":"https://www.banks.k12.ga.us/apps/news/category/19174?pageIndex=6","place_id":"65a603fbc769fa16f659736a","requestsFinished":501,"requestsFailed":0,"retryHistogram":[501],"requestAvgFailedDurationMillis":null,"requestAvgFinishedDurationMillis":6102,"requestsFinishedPerMinute":18,"requestsFailedPerMinute":0,"requestTotalDurationMillis":3057100,"requestsTotal":501,"crawlerRuntimeMillis":1697643}

bbmax

501 total requests but only 18 per minute?

bbmax

can I pay you an consulting/hourly fee to check this out with me?

PPepa J

If there are no warnings or retries in the log then it seems like pretty huge website or deoptimization in code, we used to had some bad experiences with some 3rd party libraries that provided sync blocking sorting/transforming data.

can I pay you an consulting/hourly fee to check this out with me?

Unfortunately we do not provide such a service here on discord.

I mostly advice to fill some "dangerous code" areas with logs with timestamp, so it can be determined what is taking so much time. Or you may run the Actor in headful mode with maxConcurrency: 1 locally and see if you spot anything locally.

bbmax

Are you thinking event loop lag?

PPepa J

Something like that, probably not full one, since you are getting the logs, but I don't have deep knowledge there, I know that the browsers should run in separate thread, but the CDP instructions for browser should be in single thread 🤔
So first I would discover what is causing the blocking - should be easy if you can to reproduce it with maxConcurrency: 1.

bbmax

Attachment

bbmax

any puppeteer settings you know of that would make chrome take up less?

Plain Text

  launchContext: {
      launcher: puppeteerExtra,
      useIncognitoPages: true,
      launchOptions: {
        executablePath: process.env.CHROMIUM_PATH || undefined,
        headless: 'new',
        args: [
          '--no-sandbox',
          '--disable-setuid-sandbox',
          '--aggressive-cache-discard',
          '--no-zygote',
          '--disable-cache',
          '--disable-application-cache',
          '--disable-offline-load-stale-cache',
          '--disable-gpu-shader-disk-cache',
          '--disable-gpu',
          '--media-cache-size=0',
          '--disk-cache-size=0',
          '--ignore-certificate-errors',
          '--disable-dev-shm-usage',
        ],

PPepa J

I don't think you may save many resources by this... Can you just keep the website open without processing anything? Is it also taking so much cpu?

Add a reply

Join on Discord

Apify and Crawlee Official Forum

Need serious help scaling crawlee