Re: Best practice for health check URL?
On Mon, Dec 6, 2021 at 1:25 PM JumpStart < geoff.callender.jumpst...@gmail.com> wrote: > I think you’d want to make it a configuration option, so that development > can still come up quickly, but that sounds great for production. > Yes, my mental plan is to have it configurable, probably on by default, just like production mode is. > I’ve spent some time bouncing between Dmitry and Ben’s approaches. With > the latter I simplified it with Jsoup, but there are considerable > limitations to what can be preloaded that way. > I suppose we would need to make some changes to Tapestry and Tapestry-IoC internals to offer the option to preload everything, plus maybe provide some hook for user code to preload itself (although contributions to RegistryStartup may be enough for that, but I haven't investigated yet). -- Thiago
Re: Best practice for health check URL?
Hi Geoff, I extracted all the important bits from our implementation into this gist: https://gist.github.com/dmitrygusev/486ad56174450299b94cc364d3630b28 I'd start exploring it from here: https://gist.github.com/dmitrygusev/486ad56174450299b94cc364d3630b28#file-eagerloadresourceimpl-java-L94 In the gist you'll find full implementation, but configuration will likely be different depending on your project setup. You can find some examples in `TapestryEagerLoadModule` for how you can contribute variations to the warmup, e.g.: -- different locales, or other behaviours that can prepare a request thread for serving custom responses -- an example of contributing warmup logic as one of the health checks to the Dropwizard's healthcheck registry -- contributions to the `SkipEagerLoadForPatterns` to ignore certain files from direct access during warmup -- etc. Any questions or suggestions, please let me know. Regards, Dmitry On Mon, Dec 6, 2021 at 4:24 PM JumpStart < geoff.callender.jumpst...@gmail.com> wrote: > I think you’d want to make it a configuration option, so that development > can still come up quickly, but that sounds great for production. > > I’ve spent some time bouncing between Dmitry and Ben’s approaches. With > the latter I simplified it with Jsoup, but there are considerable > limitations to what can be preloaded that way. > > Dmitry, I have tried to flesh out what you’ve done, but I think my > knowledge of Tapestry internals is holding me back. You seem to have taken > a really deep dive! Is there any code that you’d feel comfortable sharing? > > Cheers, > > Geoff > > > On 4 Dec 2021, at 6:00 am, Thiago H. de Paula Figueiredo < > thiag...@gmail.com> wrote: > > > > Hi! > > > > Today I started wondering about how we could get Tapestry to run under > > Quarkus.io, including generating a native executable. Of course, this > won't > > include bytecode generated in runtime, something many libraries and > > frameworks do, Tapestry very much included. Then I researched a bit and > > found this: > https://quarkus.io/guides/writing-extensions#bytecode-recording. > > Basically, it's a hook for you to run the code that will generate > bytecode > > while the hook records everything (if I got it right). So, to write an > > extension for Tapestry, we would need to have every page and and every > > service (and maybe some other stuff too) fully realized, since > Tapestry-IoC > > and Tapestry load mostly everything in a lazy manner. This is something > > that could also solve Geoff's question: if we can somehow force Tapestry > to > > preload everything, then the app is ready and (at least mostly) warm when > > the first request is properly served. > > > > With the code Dmitry shared here, I wonder if you want to collaborate on > > implementing this preload feature on Tapestry itself out-of-the-box, > > avoiding some ugly workarounds needed since there's no actual support for > > that. :) > > > > Cheers! > > > > On Mon, Nov 29, 2021 at 7:36 PM Dmitry Gusev > wrote: > > > >> Hi Geoff, > >> > >> I don't think there's a simpler way, we're doing something similar. > >> > >> We created a REST endpoint with tynamo-resteasy which is effectively a > load > >> balancer health check. > >> On the first hit it starts the warmup process on the same request, > >> following requests return an error instantly if the initial warmup > routine > >> is still in progress. > >> > >> Our warmup logic is heavily based on tapestry internals & reflection, in > >> conjunction with eager loading services as described here: > >> https://gist.github.com/dmitrygusev/5562739 > >> > >> Warmup logic is a bit complicated, it's trying to: > >> - "touch" each page using ComponentClassResolver.getPageNames() > >> - recursively for each component with mixins on a page starting from > root > >> component, > >>> find imported assets via reflection (fields with names starting as > >> `importedAssets_`) > >>> stream each asset via `StreamableResourceSource` into no-op consumer > >> - find JS modules with `ModuleManager` and stream through no-op consumer > >> - every JS stack returned from `JavaScriptStackSource` assemble with > >> `JavaScriptStackAssembler` and stream through no-op consumer > >> - repeat above for each locale/axis > >> > >> Entire process usually takes 3-5 minute in our setup. > >> After it's done we return 200 to the load balancer and the first real > >> request is handled with hot caches. > >> > >> Hope this helps, > >> Dmitry > >> > >> On Mon, Nov 29, 2021 at 9:53 PM JumpStart < > >> geoff.callender.jumpst...@gmail.com> wrote: > >> > >>> Any suggestions on best ways to write a “health check” page to be > called > >>> by load balancers? > >>> > >>> My app is getting big, and the traffic is big. If the app fails > >> (hopefully > >>> never, but it’s a JVM) and the traffic is heavy enough, startup never > >> seems > >>> to complete - every request times out, the app log goes quiet, and CPU > >> goes > >>> to 100%. It
Re: Best practice for health check URL?
I think you’d want to make it a configuration option, so that development can still come up quickly, but that sounds great for production. I’ve spent some time bouncing between Dmitry and Ben’s approaches. With the latter I simplified it with Jsoup, but there are considerable limitations to what can be preloaded that way. Dmitry, I have tried to flesh out what you’ve done, but I think my knowledge of Tapestry internals is holding me back. You seem to have taken a really deep dive! Is there any code that you’d feel comfortable sharing? Cheers, Geoff > On 4 Dec 2021, at 6:00 am, Thiago H. de Paula Figueiredo > wrote: > > Hi! > > Today I started wondering about how we could get Tapestry to run under > Quarkus.io, including generating a native executable. Of course, this won't > include bytecode generated in runtime, something many libraries and > frameworks do, Tapestry very much included. Then I researched a bit and > found this: https://quarkus.io/guides/writing-extensions#bytecode-recording. > Basically, it's a hook for you to run the code that will generate bytecode > while the hook records everything (if I got it right). So, to write an > extension for Tapestry, we would need to have every page and and every > service (and maybe some other stuff too) fully realized, since Tapestry-IoC > and Tapestry load mostly everything in a lazy manner. This is something > that could also solve Geoff's question: if we can somehow force Tapestry to > preload everything, then the app is ready and (at least mostly) warm when > the first request is properly served. > > With the code Dmitry shared here, I wonder if you want to collaborate on > implementing this preload feature on Tapestry itself out-of-the-box, > avoiding some ugly workarounds needed since there's no actual support for > that. :) > > Cheers! > > On Mon, Nov 29, 2021 at 7:36 PM Dmitry Gusev wrote: > >> Hi Geoff, >> >> I don't think there's a simpler way, we're doing something similar. >> >> We created a REST endpoint with tynamo-resteasy which is effectively a load >> balancer health check. >> On the first hit it starts the warmup process on the same request, >> following requests return an error instantly if the initial warmup routine >> is still in progress. >> >> Our warmup logic is heavily based on tapestry internals & reflection, in >> conjunction with eager loading services as described here: >> https://gist.github.com/dmitrygusev/5562739 >> >> Warmup logic is a bit complicated, it's trying to: >> - "touch" each page using ComponentClassResolver.getPageNames() >> - recursively for each component with mixins on a page starting from root >> component, >>> find imported assets via reflection (fields with names starting as >> `importedAssets_`) >>> stream each asset via `StreamableResourceSource` into no-op consumer >> - find JS modules with `ModuleManager` and stream through no-op consumer >> - every JS stack returned from `JavaScriptStackSource` assemble with >> `JavaScriptStackAssembler` and stream through no-op consumer >> - repeat above for each locale/axis >> >> Entire process usually takes 3-5 minute in our setup. >> After it's done we return 200 to the load balancer and the first real >> request is handled with hot caches. >> >> Hope this helps, >> Dmitry >> >> On Mon, Nov 29, 2021 at 9:53 PM JumpStart < >> geoff.callender.jumpst...@gmail.com> wrote: >> >>> Any suggestions on best ways to write a “health check” page to be called >>> by load balancers? >>> >>> My app is getting big, and the traffic is big. If the app fails >> (hopefully >>> never, but it’s a JVM) and the traffic is heavy enough, startup never >> seems >>> to complete - every request times out, the app log goes quiet, and CPU >> goes >>> to 100%. It appears to be due to race conditions possibly involving asset >>> compression, minimising, and first time into the pages. >>> >>> I’m considering having a startup service crawl every page, in every >>> language, in every skinning, before setting a singleton flag that the >>> health check page will read to determine whether the app is ready to >>> receive traffic. >>> >>> Is there a simpler way? >>> >>> Geoff >>> - >>> To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org >>> For additional commands, e-mail: users-h...@tapestry.apache.org >>> >>> >> > > > -- > Thiago - To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org For additional commands, e-mail: users-h...@tapestry.apache.org
Re: Best practice for health check URL?
Hi! Today I started wondering about how we could get Tapestry to run under Quarkus.io, including generating a native executable. Of course, this won't include bytecode generated in runtime, something many libraries and frameworks do, Tapestry very much included. Then I researched a bit and found this: https://quarkus.io/guides/writing-extensions#bytecode-recording. Basically, it's a hook for you to run the code that will generate bytecode while the hook records everything (if I got it right). So, to write an extension for Tapestry, we would need to have every page and and every service (and maybe some other stuff too) fully realized, since Tapestry-IoC and Tapestry load mostly everything in a lazy manner. This is something that could also solve Geoff's question: if we can somehow force Tapestry to preload everything, then the app is ready and (at least mostly) warm when the first request is properly served. With the code Dmitry shared here, I wonder if you want to collaborate on implementing this preload feature on Tapestry itself out-of-the-box, avoiding some ugly workarounds needed since there's no actual support for that. :) Cheers! On Mon, Nov 29, 2021 at 7:36 PM Dmitry Gusev wrote: > Hi Geoff, > > I don't think there's a simpler way, we're doing something similar. > > We created a REST endpoint with tynamo-resteasy which is effectively a load > balancer health check. > On the first hit it starts the warmup process on the same request, > following requests return an error instantly if the initial warmup routine > is still in progress. > > Our warmup logic is heavily based on tapestry internals & reflection, in > conjunction with eager loading services as described here: > https://gist.github.com/dmitrygusev/5562739 > > Warmup logic is a bit complicated, it's trying to: > - "touch" each page using ComponentClassResolver.getPageNames() > - recursively for each component with mixins on a page starting from root > component, > > find imported assets via reflection (fields with names starting as > `importedAssets_`) > > stream each asset via `StreamableResourceSource` into no-op consumer > - find JS modules with `ModuleManager` and stream through no-op consumer > - every JS stack returned from `JavaScriptStackSource` assemble with > `JavaScriptStackAssembler` and stream through no-op consumer > - repeat above for each locale/axis > > Entire process usually takes 3-5 minute in our setup. > After it's done we return 200 to the load balancer and the first real > request is handled with hot caches. > > Hope this helps, > Dmitry > > On Mon, Nov 29, 2021 at 9:53 PM JumpStart < > geoff.callender.jumpst...@gmail.com> wrote: > > > Any suggestions on best ways to write a “health check” page to be called > > by load balancers? > > > > My app is getting big, and the traffic is big. If the app fails > (hopefully > > never, but it’s a JVM) and the traffic is heavy enough, startup never > seems > > to complete - every request times out, the app log goes quiet, and CPU > goes > > to 100%. It appears to be due to race conditions possibly involving asset > > compression, minimising, and first time into the pages. > > > > I’m considering having a startup service crawl every page, in every > > language, in every skinning, before setting a singleton flag that the > > health check page will read to determine whether the app is ready to > > receive traffic. > > > > Is there a simpler way? > > > > Geoff > > - > > To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org > > For additional commands, e-mail: users-h...@tapestry.apache.org > > > > > -- Thiago
Re: Best practice for health check URL?
Hi Ben, This is all truly great to know. I think I may draw ideas from both yours and Dmitry’s response. Thank you, Geoff > On 30 Nov 2021, at 4:40 pm, Ben Weidig wrote: > > Hi Geoff, > > we have a multi-tenant/-domain multi-language site that required pre-heated > servers after deployment, so we created a WarmupTaskRunner to run at > startup, which gets contributed different kinds of tasks. > They can either be blocking or just run in the background, depending on > what they do (e.g., pre-heating caches/assets, ensuring indices in MongoDB > are set in dynamic collections, ...). > > In our case, we just call the index pages for all relevant domains with > Apache HttpClient instead of creating the pages and components in Tapestry > itself. > But Dmitry's way is more thoroughly than our method. > > The WarmupTaskRunner is combined with the health checks for HAProxy. > An HttpServletFilter checks for things like "warmup finished" by asking the > WarmupTaskRunner, and "mysql connection pool available", "mongodb > available", etc., and returns an appropriate HTTP status code to HAProxy. > If not 200 is returned, HAProxy takes the server out of rotation. > > Cheers, > Ben > > > On Tue, Nov 30, 2021 at 2:41 AM JumpStart < > geoff.callender.jumpst...@gmail.com> wrote: > >> Hi Dmitry, >> >> That is spectacularly helpful! >> >> We’re about to write a headless smoke test anyway that will visit every >> page. Do you see any downside to using that to do the warmup? >> >> Where do you keep your shared “warmup in progress” flag so that it is >> rapidly accessible on every health check request? >> >> Cheers, >> >> Geoff >> >>> On 30 Nov 2021, at 6:35 am, Dmitry Gusev wrote: >>> >>> Hi Geoff, >>> >>> I don't think there's a simpler way, we're doing something similar. >>> >>> We created a REST endpoint with tynamo-resteasy which is effectively a >> load >>> balancer health check. >>> On the first hit it starts the warmup process on the same request, >>> following requests return an error instantly if the initial warmup >> routine >>> is still in progress. >>> >>> Our warmup logic is heavily based on tapestry internals & reflection, in >>> conjunction with eager loading services as described here: >>> https://gist.github.com/dmitrygusev/5562739 >>> >>> Warmup logic is a bit complicated, it's trying to: >>> - "touch" each page using ComponentClassResolver.getPageNames() >>> - recursively for each component with mixins on a page starting from root >>> component, find imported assets via reflection (fields with names starting as >>> `importedAssets_`) stream each asset via `StreamableResourceSource` into no-op consumer >>> - find JS modules with `ModuleManager` and stream through no-op consumer >>> - every JS stack returned from `JavaScriptStackSource` assemble with >>> `JavaScriptStackAssembler` and stream through no-op consumer >>> - repeat above for each locale/axis >>> >>> Entire process usually takes 3-5 minute in our setup. >>> After it's done we return 200 to the load balancer and the first real >>> request is handled with hot caches. >>> >>> Hope this helps, >>> Dmitry >>> >>> On Mon, Nov 29, 2021 at 9:53 PM JumpStart < >>> geoff.callender.jumpst...@gmail.com> wrote: >>> Any suggestions on best ways to write a “health check” page to be called by load balancers? My app is getting big, and the traffic is big. If the app fails >> (hopefully never, but it’s a JVM) and the traffic is heavy enough, startup never >> seems to complete - every request times out, the app log goes quiet, and CPU >> goes to 100%. It appears to be due to race conditions possibly involving >> asset compression, minimising, and first time into the pages. I’m considering having a startup service crawl every page, in every language, in every skinning, before setting a singleton flag that the health check page will read to determine whether the app is ready to receive traffic. Is there a simpler way? Geoff - To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org For additional commands, e-mail: users-h...@tapestry.apache.org >> >> >> - >> To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org >> For additional commands, e-mail: users-h...@tapestry.apache.org >> >> - To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org For additional commands, e-mail: users-h...@tapestry.apache.org
Re: Best practice for health check URL?
Hi Geoff, RESTEasy services are singletons, so we simply created an `AtomicReference` field and used `compareAndSet()`, something like this: private final AtomicReference warmUpStatus = new AtomicReference<>(WarmUpStatus.EMPTY); @Override public Response warmUp() { if (warmUpStatus.compareAndSet(WarmUpStatus.EMPTY, WarmUpStatus.IN_PROGRESS)) { Regards, Dmitry On Tue, Nov 30, 2021 at 1:41 AM JumpStart < geoff.callender.jumpst...@gmail.com> wrote: > Hi Dmitry, > > That is spectacularly helpful! > > We’re about to write a headless smoke test anyway that will visit every > page. Do you see any downside to using that to do the warmup? > > Where do you keep your shared “warmup in progress” flag so that it is > rapidly accessible on every health check request? > > Cheers, > > Geoff > > > On 30 Nov 2021, at 6:35 am, Dmitry Gusev wrote: > > > > Hi Geoff, > > > > I don't think there's a simpler way, we're doing something similar. > > > > We created a REST endpoint with tynamo-resteasy which is effectively a > load > > balancer health check. > > On the first hit it starts the warmup process on the same request, > > following requests return an error instantly if the initial warmup > routine > > is still in progress. > > > > Our warmup logic is heavily based on tapestry internals & reflection, in > > conjunction with eager loading services as described here: > > https://gist.github.com/dmitrygusev/5562739 > > > > Warmup logic is a bit complicated, it's trying to: > > - "touch" each page using ComponentClassResolver.getPageNames() > > - recursively for each component with mixins on a page starting from root > > component, > >> find imported assets via reflection (fields with names starting as > > `importedAssets_`) > >> stream each asset via `StreamableResourceSource` into no-op consumer > > - find JS modules with `ModuleManager` and stream through no-op consumer > > - every JS stack returned from `JavaScriptStackSource` assemble with > > `JavaScriptStackAssembler` and stream through no-op consumer > > - repeat above for each locale/axis > > > > Entire process usually takes 3-5 minute in our setup. > > After it's done we return 200 to the load balancer and the first real > > request is handled with hot caches. > > > > Hope this helps, > > Dmitry > > > > On Mon, Nov 29, 2021 at 9:53 PM JumpStart < > > geoff.callender.jumpst...@gmail.com> wrote: > > > >> Any suggestions on best ways to write a “health check” page to be called > >> by load balancers? > >> > >> My app is getting big, and the traffic is big. If the app fails > (hopefully > >> never, but it’s a JVM) and the traffic is heavy enough, startup never > seems > >> to complete - every request times out, the app log goes quiet, and CPU > goes > >> to 100%. It appears to be due to race conditions possibly involving > asset > >> compression, minimising, and first time into the pages. > >> > >> I’m considering having a startup service crawl every page, in every > >> language, in every skinning, before setting a singleton flag that the > >> health check page will read to determine whether the app is ready to > >> receive traffic. > >> > >> Is there a simpler way? > >> > >> Geoff > >> - > >> To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org > >> For additional commands, e-mail: users-h...@tapestry.apache.org > >> > >> > > > - > To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org > For additional commands, e-mail: users-h...@tapestry.apache.org > > -- Dmitry Gusev AnjLab Team http://anjlab.com
Re: Best practice for health check URL?
Hi Geoff, we have a multi-tenant/-domain multi-language site that required pre-heated servers after deployment, so we created a WarmupTaskRunner to run at startup, which gets contributed different kinds of tasks. They can either be blocking or just run in the background, depending on what they do (e.g., pre-heating caches/assets, ensuring indices in MongoDB are set in dynamic collections, ...). In our case, we just call the index pages for all relevant domains with Apache HttpClient instead of creating the pages and components in Tapestry itself. But Dmitry's way is more thoroughly than our method. The WarmupTaskRunner is combined with the health checks for HAProxy. An HttpServletFilter checks for things like "warmup finished" by asking the WarmupTaskRunner, and "mysql connection pool available", "mongodb available", etc., and returns an appropriate HTTP status code to HAProxy. If not 200 is returned, HAProxy takes the server out of rotation. Cheers, Ben On Tue, Nov 30, 2021 at 2:41 AM JumpStart < geoff.callender.jumpst...@gmail.com> wrote: > Hi Dmitry, > > That is spectacularly helpful! > > We’re about to write a headless smoke test anyway that will visit every > page. Do you see any downside to using that to do the warmup? > > Where do you keep your shared “warmup in progress” flag so that it is > rapidly accessible on every health check request? > > Cheers, > > Geoff > > > On 30 Nov 2021, at 6:35 am, Dmitry Gusev wrote: > > > > Hi Geoff, > > > > I don't think there's a simpler way, we're doing something similar. > > > > We created a REST endpoint with tynamo-resteasy which is effectively a > load > > balancer health check. > > On the first hit it starts the warmup process on the same request, > > following requests return an error instantly if the initial warmup > routine > > is still in progress. > > > > Our warmup logic is heavily based on tapestry internals & reflection, in > > conjunction with eager loading services as described here: > > https://gist.github.com/dmitrygusev/5562739 > > > > Warmup logic is a bit complicated, it's trying to: > > - "touch" each page using ComponentClassResolver.getPageNames() > > - recursively for each component with mixins on a page starting from root > > component, > >> find imported assets via reflection (fields with names starting as > > `importedAssets_`) > >> stream each asset via `StreamableResourceSource` into no-op consumer > > - find JS modules with `ModuleManager` and stream through no-op consumer > > - every JS stack returned from `JavaScriptStackSource` assemble with > > `JavaScriptStackAssembler` and stream through no-op consumer > > - repeat above for each locale/axis > > > > Entire process usually takes 3-5 minute in our setup. > > After it's done we return 200 to the load balancer and the first real > > request is handled with hot caches. > > > > Hope this helps, > > Dmitry > > > > On Mon, Nov 29, 2021 at 9:53 PM JumpStart < > > geoff.callender.jumpst...@gmail.com> wrote: > > > >> Any suggestions on best ways to write a “health check” page to be called > >> by load balancers? > >> > >> My app is getting big, and the traffic is big. If the app fails > (hopefully > >> never, but it’s a JVM) and the traffic is heavy enough, startup never > seems > >> to complete - every request times out, the app log goes quiet, and CPU > goes > >> to 100%. It appears to be due to race conditions possibly involving > asset > >> compression, minimising, and first time into the pages. > >> > >> I’m considering having a startup service crawl every page, in every > >> language, in every skinning, before setting a singleton flag that the > >> health check page will read to determine whether the app is ready to > >> receive traffic. > >> > >> Is there a simpler way? > >> > >> Geoff > >> - > >> To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org > >> For additional commands, e-mail: users-h...@tapestry.apache.org > >> > >> > > > - > To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org > For additional commands, e-mail: users-h...@tapestry.apache.org > >
Re: Best practice for health check URL?
Hi Dmitry, That is spectacularly helpful! We’re about to write a headless smoke test anyway that will visit every page. Do you see any downside to using that to do the warmup? Where do you keep your shared “warmup in progress” flag so that it is rapidly accessible on every health check request? Cheers, Geoff > On 30 Nov 2021, at 6:35 am, Dmitry Gusev wrote: > > Hi Geoff, > > I don't think there's a simpler way, we're doing something similar. > > We created a REST endpoint with tynamo-resteasy which is effectively a load > balancer health check. > On the first hit it starts the warmup process on the same request, > following requests return an error instantly if the initial warmup routine > is still in progress. > > Our warmup logic is heavily based on tapestry internals & reflection, in > conjunction with eager loading services as described here: > https://gist.github.com/dmitrygusev/5562739 > > Warmup logic is a bit complicated, it's trying to: > - "touch" each page using ComponentClassResolver.getPageNames() > - recursively for each component with mixins on a page starting from root > component, >> find imported assets via reflection (fields with names starting as > `importedAssets_`) >> stream each asset via `StreamableResourceSource` into no-op consumer > - find JS modules with `ModuleManager` and stream through no-op consumer > - every JS stack returned from `JavaScriptStackSource` assemble with > `JavaScriptStackAssembler` and stream through no-op consumer > - repeat above for each locale/axis > > Entire process usually takes 3-5 minute in our setup. > After it's done we return 200 to the load balancer and the first real > request is handled with hot caches. > > Hope this helps, > Dmitry > > On Mon, Nov 29, 2021 at 9:53 PM JumpStart < > geoff.callender.jumpst...@gmail.com> wrote: > >> Any suggestions on best ways to write a “health check” page to be called >> by load balancers? >> >> My app is getting big, and the traffic is big. If the app fails (hopefully >> never, but it’s a JVM) and the traffic is heavy enough, startup never seems >> to complete - every request times out, the app log goes quiet, and CPU goes >> to 100%. It appears to be due to race conditions possibly involving asset >> compression, minimising, and first time into the pages. >> >> I’m considering having a startup service crawl every page, in every >> language, in every skinning, before setting a singleton flag that the >> health check page will read to determine whether the app is ready to >> receive traffic. >> >> Is there a simpler way? >> >> Geoff >> - >> To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org >> For additional commands, e-mail: users-h...@tapestry.apache.org >> >> - To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org For additional commands, e-mail: users-h...@tapestry.apache.org
Re: Best practice for health check URL?
Hi Geoff, I don't think there's a simpler way, we're doing something similar. We created a REST endpoint with tynamo-resteasy which is effectively a load balancer health check. On the first hit it starts the warmup process on the same request, following requests return an error instantly if the initial warmup routine is still in progress. Our warmup logic is heavily based on tapestry internals & reflection, in conjunction with eager loading services as described here: https://gist.github.com/dmitrygusev/5562739 Warmup logic is a bit complicated, it's trying to: - "touch" each page using ComponentClassResolver.getPageNames() - recursively for each component with mixins on a page starting from root component, > find imported assets via reflection (fields with names starting as `importedAssets_`) > stream each asset via `StreamableResourceSource` into no-op consumer - find JS modules with `ModuleManager` and stream through no-op consumer - every JS stack returned from `JavaScriptStackSource` assemble with `JavaScriptStackAssembler` and stream through no-op consumer - repeat above for each locale/axis Entire process usually takes 3-5 minute in our setup. After it's done we return 200 to the load balancer and the first real request is handled with hot caches. Hope this helps, Dmitry On Mon, Nov 29, 2021 at 9:53 PM JumpStart < geoff.callender.jumpst...@gmail.com> wrote: > Any suggestions on best ways to write a “health check” page to be called > by load balancers? > > My app is getting big, and the traffic is big. If the app fails (hopefully > never, but it’s a JVM) and the traffic is heavy enough, startup never seems > to complete - every request times out, the app log goes quiet, and CPU goes > to 100%. It appears to be due to race conditions possibly involving asset > compression, minimising, and first time into the pages. > > I’m considering having a startup service crawl every page, in every > language, in every skinning, before setting a singleton flag that the > health check page will read to determine whether the app is ready to > receive traffic. > > Is there a simpler way? > > Geoff > - > To unsubscribe, e-mail: users-unsubscr...@tapestry.apache.org > For additional commands, e-mail: users-h...@tapestry.apache.org > >