Started a new wiki page to discuss this here https://cwiki.apache.org/confluence/display/TC/Configuration+Management
I will do my best to summarize the discussion below later today. Ryan M | 303-524-5099 -----Original Message----- From: Eric Friedrich (efriedri) [mailto:[email protected]] Sent: Tuesday, April 11, 2017 8:55 AM To: [email protected] Subject: Re: Proposal for CDN definition file based configuration management A few questions/thoughts, apologies for not in-lining: 1) If we move away from individually queued updates, we give up the ability to make changes and then selectively deploy them. How often do TC operations teams make config changes but do not immediately queue updates. (I personally think that we currently have a bit of a tricky situation where queuing updates much later can push down an unknowingly large config change to a cache- i.e. many new DS added/removed since last time updates were queued maybe months earlier). I wouldn't be sad to see queue updates go away, but don't want to cause hardship on operators using that feature. 2) If we move away from individually queued updates, how does that affect the implicit "config state machine"? Specifically, how will edges know when their parents have been configured and are ready for service? Today we don't config an edge cache with a new DS unless the mid is ready to handle traffic as well. 3) If we move away from individually queued updates, how do we do things like unassign a delivery service from a cache? Today we have to snapshot CRConfig first to stop redirects to the cache before we queue the update. If updates are immediately applied and snapshot is still separate, how do we get TR to stop sending traffic to a cache that no longer has the remap rule? 4) Also along the lines of the config state machine, we never really closed on if we would make any changes to the queue update/snapshot CRConfig flow. If we are looking at redoing how we generate config files, it would be great to have consensus on an approach (if not an implementation) to remove the need to sequence queue updates and snapshot CRConfig. I think the requirement here would be to have Traffic Control figure out on its own when to activate/deactivate routing to a cache from TR. 5) I like the suggestion of cache-based config file generation. - Caches only retrieve relevant information, so scale proportional to number of caches/DSs in the CDN is much better - We could modify TR/TM to use the same approach, rather than snapshotting a CRConfig. - Cache/TR/TM-based config could play a greater role in config state machine, rather than having Traffic Ops build static configuration ahead of time. Downsides - Versioning is still possible, but more work than maintaining snapshots of a config file - Have to be very careful with API changes, any breakage now impacts cache updates. -Eric > On Apr 10, 2017, at 9:45 PM, Gelinas, Derek <[email protected]> wrote: > > Thanks Rob. To your point about scalability: I think that this is more > scaleable than the current crconfig implementation due to the caching. > However that is a very valid point and one that has been considered. I've > started looking into the problem from that angle and hope to have some more > solid data soon. I still believe that this is ultimately more scaleable than > current config implementation, even with the scope caching, but the proof > will be in the data. > > Derek > >> On Apr 10, 2017, at 9:23 PM, Robert Butts <[email protected]> wrote: >> >> I'd propose: >> * Instead of storing the JSON as blob, use >> https://www.postgresql.org/doc s/9.2/static/datatype-json.html >> * Instead of version-then-file request, use a "latest" endpoint with >> `If-Modified-Since` >> (https://tools.ietf.org/html/rfc7232#section-3.3). We can also serve >> each version at endpoints, but `If-Modified-Since` lets us determine >> whether there's a new snapshot and get it in a single request, both >> efficiently and using a standard. (We should do the same for the CRConfig). >> >> Also for cache-side config generation, consider >> https://github.com/apache/incubator-trafficcontrol/pull/151 . It's a >> prototype and needs work to bring it to production, but the basic >> functionality is there. Go is safer and faster to develop than Perl, >> and this is already encapsulated in a library, with both CLI and HTTP >> microservice examples. I'm certainly willing to help bring it to production. >> >> >> "a single definition file for each CDN which will contain all the >> information required for any server within that CDN to generate its >> own configs" >> >> Also, long-term, that doesn't scale, nor does the CRConfig. As >> Traffic Control is deployed with larger and larger CDNs, the CRConfig >> grows uncontrollably. It's already 5-7mb for us, which takes an >> approaching-unreasonable amount of time for Traffic Monitor and >> Router to fetch. This isn't an immediate concern, but long-term, we >> need to develop a scalable solution, something that says "only give >> me the data modified since this timestamp". >> >> Again, this isn't an immediate crisis. I only mention it now because, >> if a scalable solution is about the same amount of work, now sounds >> like a good time. If it's relevantly more work, no worries. >> >> >> But otherwise, +1. We've long needed to Separate our Concerns of >> Traffic Ops and the cache application. >> >> >> On Mon, Apr 10, 2017 at 5:05 PM, Gelinas, Derek >> <[email protected]> >> wrote: >> >>> I would like to propose a new method for ATS config file generation, >>> in which a single definition file for each CDN which will contain >>> all the information required for any server within that CDN to >>> generate its own configs, rather than requesting them from traffic >>> ops. This would be a version-controlled json file that, when >>> generated, would be stored in a new table in the traffic ops >>> database as a blob type. This will satisfy high-availability >>> requirements and allow several versions of the configuration to be >>> retained for rollback, as well as "freezing" the config at that >>> moment in time. Combined with cache support coming in 2.1, this file would >>> only need be generated once per traffic ops server instance. >>> Instead of queueing servers to update their configurations, the >>> configuration would be snapshotted similar to the crconfig file and >>> downloaded by each cache according to their set interval checks - >>> rather than performing a syncds and checking that the server has >>> been queued for update, the version number would simply be checked >>> and compared against the currently active version on the cache >>> itself. Should a difference be found the server would request the >>> definition file and begin generating configuration files for itself using >>> the data in the definition file. >>> >>> I would like feedback from the community regarding this proposal, >>> and any suggestions or comments you may have. >>> >>> Thanks! >>> Derek >>> >>> Derek Gelinas >>> IPCDN Engineering >>> [email protected]<mailto:[email protected] >>> om> >>> 603.812.5379 >>> >>>
