LGTM > The biggest advantage of ALS, is replacing the physical local access log file.
The APISIX has a plugin `http-logger`, I think APISIX is easy to report those data to SkyWalking. And I think we can try to use `gRPC` to report data. What do you think? @ming @spacewander On Fri, Mar 19, 2021 at 11:43 PM Ming Wen <wenm...@apache.org> wrote: > Cool, I think we can try ALS > > Sheng Wu <wu.sheng.841...@gmail.com>于2021年3月19日 周五下午11:21写道： > > > We have run the Envoy ALS performance test a long time ago(2 years, I > > think). Sadly, no public blog. > > It could easily support 10k-20k/s traffic with a very low load to the > > proxy. But, it is based on gRPC, so basically, it doesn't send many > things > > out, mostly only flags and latency value. > > APISIX is using HTTP1.1 mostly, so, this could be different. > > Also, at the same time, besides the ALS, we have a chance to provide a > push > > mode LUA SDK in skywalking meter format. It could provide URI-level > metrics > > at reasonable resource cost, especially it works with SkyWalking > satellite, > > which could leverage local disk to cache the data. > > The biggest advantage of ALS, is replacing the physical local access log > > file. So you have metrics, topology(dependency), and logs in one package > of > > data. > > > > If you want to discuss most ALS in APISIX, I am glad to help with the > > design and integration solution. > > > > Sheng Wu 吴晟 > > Twitter, wusheng1108 > > > > > > Ming Wen <wenm...@apache.org> 于2021年3月19日周五 下午11:08写道： > > > > > yes, Prometheus is not designed for URI-level metrics. > > > > > > Does envoy ALS will send too much data and cause performance problems? > > > > > > And I know skywalking will supoort envoy WASM, is there anything apisix > > can > > > learn from this？ > > > > > > Sheng Wu <wu.sheng.841...@gmail.com>于2021年3月19日 周五下午8:11写道： > > > > > > > Hi APISIX & SkyWalking community > > > > > > > > I was discussing with Wei Jin about using Prometheus SDK to monitor > > > > URI-level metrics of APISIX/OpenResty. > > > > Wei received feedback that way of using Prometheus SDK causing > > > > a significant performance impact. From the observability perspective, > > > Prom > > > > is not good at gathering a large scale of entities' metrics due to > its > > > > local cache and pull mode. > > > > Instead, we should consider providing a push way like Envoy ALS[1] > and > > > > Metrics Service[2] did(it does through gRPC, and SkyWalking > recommended > > > in > > > > the prod environment. > > > > If this is accepted and agreed by the APISIX community, we could look > > for > > > > implementation in the skywalking-nginx-lua[3] repo to collect > > > metrics(meter > > > > in SkyWalking) and use HTTP(in SkyWalking native meter format) to > > report > > > > > > > > This mail has been sent to dev@apisix.apache.org and > > > > d...@skywalking.apache.org. Please reply all when you discuss in the > > > > thread. > > > > > > > > [1] > > > > > > > > > > > > > > https://www.envoyproxy.io/docs/envoy/latest/api-v2/config/accesslog/v2/als.proto > > > > [2] > > > > > > > > > > > > > > https://www.envoyproxy.io/docs/envoy/latest/api-v2/config/metrics/v2/metrics_service.proto > > > > [3] https://github.com/apache/skywalking-nginx-lua > > > > > > > > Sheng Wu 吴晟 > > > > Twitter, wusheng1108 > > > > > > > -- > > > Thanks, > > > Ming Wen, Apache APISIX PMC Chair > > > Twitter: _WenMing > > > > > > -- > Thanks, > Ming Wen, Apache APISIX PMC Chair > Twitter: _WenMing > -- *MembPhis* My GitHub: https://github.com/membphis Apache APISIX: https://github.com/apache/apisix
