[ http://issues.apache.org/jira/browse/NUTCH-385?page=comments#action_12441528 ] Chris Schneider commented on NUTCH-385: ---------------------------------------
This comment was actually made by Andrzej in response to an email containing the analysis above that I sent him before creating this JIRA issue: Let's start with defining what is the desired semantics of these two parameters together. In my opinion it's the following: * if only 1 thread per host is allowed, at any given moment at most one thread should be accessing the host, and the interval between consecutive requests should be at least crawlDelay (whichever way we determine this value - from config, from robots.txt or external sources such as partner agreements). * if two or more (for example N) threads per host are allowed, at any given moment at most N threads should be accessing the host, and the interval between consecutive requests should be at least crawlDelay - that is, the interval between when one of the threads finishes, and another starts requesting. I.e.: for threads.per.host=2 and crawlDelay=3 seconds, if we start 3 threads trying to access the same host we should get something like this (time in [s] on the x axis, # - start request, + - request in progress, b - blocked in per-host limit, c - obeying crawlDelay): ===0 1 2 ===01234567890123456789012345678 1: #+++cccbbccc#++++cccbb#++++++ 2: #++++++++cccbcccbcc#+++cccccb 3: bbbbccc#+++++ccc#+++++ccc#+++ As you can see, at any given time we have at most 2 threads accessing the site, and the interval between consecutive requests is at least 3 seconds. Especially interesting in the above graph is the period between 17-18 seconds - thread 2 had to be delayed additional 2 seconds to satisfy the crawl delay requirement, even though the threads.per.host requirement was satisfied. [snip] It's a question of priorities - in the model I drafted above the topmost priority is the observance of crawlDelay, sometimes at the cost of the number of concurrent threads (see seconds 17-18). In this model, the code should always put the delay in BLOCKED_ADDR_TO_TIME, in order to wait at least crawlDelay after _any_ thread finishes. We could use an alternative model, where crawlDelay is measured from the start of the request, and not from the end - see the graph below: ===0 1 2 3 ===01234567890123456789012345678901234567 1: #+++cccccbbb#++++cccc#++++++cc#+++++++ 2: ccc#++++++++cccccc#+++ccccc#++++c#++++ 3: cccccc#+++++ccc#+++++ccc#+++++ccccccbb but it seems to me that it's more complicated, gives less requests/sec, and the interpretaion of crawlDelay's meaning is stretched ... [snip] > Server delay feature conflicts with maxThreadsPerHost > ----------------------------------------------------- > > Key: NUTCH-385 > URL: http://issues.apache.org/jira/browse/NUTCH-385 > Project: Nutch > Issue Type: Bug > Components: fetcher > Reporter: Chris Schneider > > For some time I've been puzzled by the interaction between two paramters that > control how often the fetcher can access a particular host: > 1) The server delay, which comes back from the remote server during our > processing of the robots.txt file, and which can be limited by > fetcher.max.crawl.delay. > 2) The fetcher.threads.per.host value, particularly when this is greater than > the default of 1. > According to my (limited) understanding of the code in HttpBase.java: > Suppose that fetcher.threads.per.host is 2, and that (by chance) the fetcher > ends up keeping either 1 or 2 fetcher threads pointing at a particular host > continuously. In other words, it never tries to point 3 at the host, and it > always points a second thread at the host before the first thread finishes > accessing it. Since HttpBase.unblockAddr never gets called with > (((Integer)THREADS_PER_HOST_COUNT.get(host)).intValue() == 1), it never puts > System.currentTimeMillis() + crawlDelay into BLOCKED_ADDR_TO_TIME for the > host. Thus, the server delay will never be used at all. The fetcher will be > continuously retrieving pages from the host, often with 2 fetchers accessing > the host simultaneously. > Suppose instead that the fetcher finally does allow the last thread to > complete before it gets around to pointing another thread at the target host. > When the last fetcher thread calls HttpBase.unblockAddr, it will now put > System.currentTimeMillis() + crawlDelay into BLOCKED_ADDR_TO_TIME for the > host. This, in turn, will prevent any threads from accessing this host until > the delay is complete, even though zero threads are currently accessing the > host. > I see this behavior as inconsistent. More importantly, the current > implementation certainly doesn't seem to answer my original question about > appropriate definitions for what appear to be conflicting parameters. > In a nutshell, how could we possibly honor the server delay if we allow more > than one fetcher thread to simultaneously access the host? > It would be one thing if whenever (fetcher.threads.per.host > 1), this > trumped the server delay, causing the latter to be ignored completely. That > is certainly not the case in the current implementation, as it will wait for > server delay whenever the number of threads accessing a given host drops to > zero. -- This message is automatically generated by JIRA. - If you think it was sent incorrectly contact one of the administrators: http://issues.apache.org/jira/secure/Administrators.jspa - For more information on JIRA, see: http://www.atlassian.com/software/jira
