devtrackers.gg
almost 2 years
ago -
EVE Online
-
Direct link
Transcript (by Youtube)
| 1s | uh hi guys so my name is nick herring |
|---|---|
| 3s | i'm the technical director of |
| 4s | infrastructure for eve online at ccp |
| 6s | games which is a very long title for |
| 8s | cosmic plumber um we're going to be |
| 10s | going over kind of eve online |
| 13s | the last 20 years of development what |
| 15s | that looks like what we had to work on |
| 17s | to kind of modernize the eve online |
| 20s | tech stack at least from the perspective |
| 21s | of the server side and the network |
| 23s | portions of the client side |
| 25s | um |
| 27s | and we'll go over more of what the |
| 29s | topology is in the evolution of that |
| 31s | topology kind of how you've originally |
| 33s | started and how it's gone from |
| 35s | uh |
| 36s | from 2003 to what we have now |
| 38s | and |
| 39s | we're going to talk about how we tried |
| 41s | to fundamentally or are fundamentally |
| 43s | changing how we actually work on eve |
| 45s | online |
| 46s | there's multiple pieces to that but the |
| 48s | two biggest ones being a technical |
| 49s | aspect and a cultural aspect |
| 51s | and the cultural aspect is a pretty big |
| 53s | part of it |
| 55s | and hopefully |
| 57s | we don't have to go too fast here |
| 58s | because right after this we have a round |
| 59s | table um but the round table is more for |
| 62s | anything else so any kind of quasar |
| 64s | specific stuff we can talk about here |
| 65s | hopefully at the end of this if there's |
| 67s | time for questions uh and then |
| 70s | afterwards in the round type we can talk |
| 71s | more about other things like quasar and |
| 73s | easy and how they interact and what |
| 74s | makes sense there and the any kind of |
| 77s | other technology that we're using on the |
| 78s | server side |
| 80s | so we can start with |
| 81s | 20 years of eve development uh it was |
| 84s | released in 2003. you guys all know this |
| 88s | right now there are roughly over 2 |
| 90s | million changeless in perforce |
| 92s | that number is probably growing faster |
| 94s | and faster as we add more and more |
| 96s | automation into the ecosystem so there's |
| 98s | less and less humans actually making |
| 99s | changes to the code base |
| 102s | and |
| 102s | we've kind of added a little reference |
| 105s | of how much code there is and i've added |
| 107s | this silly reference of the ue4s so if |
| 110s | you take the code base of unreal engine |
| 112s | 4 |
| 112s | you can kind of get an idea of how much |
| 115s | of that code is is being used there it |
| 117s | means absolutely nothing it's just fun |
| 118s | to think about |
| 119s | um |
| 121s | and so if we think about like the cec |
| 122s | plus plus that's where a lot of the |
| 124s | rendering code is that's where the |
| 126s | simulation code is |
| 127s | and that's a lot of where the the glue |
| 129s | is from uh |
| 131s | c to python marshalling and back and |
| 133s | forth |
| 134s | so that's roughly 1.7 million lines of |
| 137s | code |
| 138s | then the next up would be sql so a lot |
| 141s | of eve is run by basically sql procs a |
| 144s | lot of the logic is |
| 146s | unfortunately um |
| 149s | and so we can see that |
| 150s | just our sql code alone is the size of |
| 153s | ue4 |
| 156s | and then if we continue on |
| 158s | where a bulk of the logic was written uh |
| 161s | in python in stackless python |
| 163s | um we can see that there's 3.4 million |
| 165s | lines of code a little bit more than a |
| 167s | ue4 |
| 169s | and then if you weren't worried enough |
| 170s | yet |
| 171s | we have roughly |
| 173s | 53 million lines of yaml |
| 176s | um |
| 177s | this is 24 unreal engines uh worth of |
| 180s | code code's a strong word but um |
| 185s | this this is what holds everything to do |
| 187s | with uh how the universe is authored any |
| 190s | anywhere from how the spaceships are |
| 192s | made uh and authored as far the |
| 194s | attributes are concerned and there's a |
| 196s | lot of work done there and and this this |
| 198s | would have looked a lot differently uh |
| 200s | from |
| 201s | probably about three years ago i think |
| 203s | where we had a team internally that was |
| 205s | jamming on moving all of that out of sql |
| 208s | into a binary file system that we could |
| 210s | work with an author with that way we |
| 211s | didn't have to deal with things like |
| 213s | promoting between databases and instead |
| 215s | the files traveled with the actual |
| 217s | branch itself |
| 219s | so that's just a little bit of idea of |
| 220s | kind of |
| 221s | the momentum that is eve online and what |
| 224s | we have to deal with when we want to |
| 226s | make a foundational change to how the |
| 228s | system works |
| 230s | and so looking at the beginning here |
| 232s | it is deceptively simple |
| 234s | uh most of you guys know about this i've |
| 236s | seen this in a different form |
| 238s | uh we have the concept of soul nodes |
| 242s | that is part of the monolithic code base |
| 243s | that we have soul nodes can take on any |
| 247s | role as it were it could be servicing |
| 250s | corporation requests alliance requests |
| 253s | wallet uh the actual literal location of |
| 256s | the solar system the simulation within |
| 257s | that so we can swap in and out of those |
| 259s | roles and a lot of the orchestration |
| 260s | usually happens on the soul node level |
| 262s | uh then we also have the proxies which |
| 264s | kind of dedicated to |
| 267s | managing the connections coming into the |
| 269s | system and that's important because |
| 271s | ultimately that wasn't the like |
| 273s | that wasn't the original version of this |
| 276s | and and we have some other ccps here who |
| 279s | actually just rejoined that were in the |
| 280s | original team that worked on this they |
| 281s | can probably pick this apart a little |
| 282s | bit more but proxies didn't originally |
| 284s | exist they were built out of necessity |
| 287s | basically as soon as people went over |
| 288s | 100 people on eve they were like we have |
| 290s | to do something about how this connects |
| 292s | because ultimately |
| 295s | this represents kind of the the topology |
| 297s | of carbon io so everything is |
| 300s | is a mesh network so it's a guaranteed |
| 302s | one-hop mesh network which becomes a |
| 304s | quadratic problem almost immediately |
| 307s | when you're trying to deal with resource |
| 308s | management and those kind of things but |
| 310s | it's very powerful in the sense that a |
| 312s | lot of how eve works is about |
| 313s | deterministic routing |
| 316s | that's contrary to modern technologies |
| 318s | around things like rabbitmq or nats or |
| 321s | kafka those types of things where you |
| 322s | have more dynamic routing slightly |
| 324s | different but |
| 325s | eve does a lot of things where it knows |
| 327s | your character id |
| 328s | and by virtue of knowing your character |
| 330s | id it doesn't need to ask the ecosystem |
| 332s | where to go it can make a very educated |
| 335s | guess and get to the right node for the |
| 337s | right information and that's very |
| 338s | powerful early on |
| 340s | and all of this was kind of |
| 343s | tackled around |
| 345s | carbon io which is kind of the glue for |
| 346s | all of it at this point in time |
| 348s | and this is a homebrew protocol written |
| 350s | in in python so a lot of the networking |
| 352s | calls are all pure python and that's |
| 354s | where all the traffic is going and being |
| 355s | shaped and those kind of things |
| 357s | and that goes back into the desktop |
| 358s | client over the internet and those types |
| 359s | of things |
| 360s | another big part of this is the |
| 362s | combination of this with with i o |
| 364s | completion ports |
| 365s | and io completion ports are important in |
| 367s | this regard |
| 368s | because with stackless python stackless |
| 370s | python only does one thing at a time |
| 373s | and it's something that i have to remind |
| 374s | every engineer it can only do one thing |
| 377s | at a time |
| 378s | engineers try to do modern techniques of |
| 380s | like distributed locks or mutexes |
| 382s | whatever the case may be but it doesn't |
| 383s | actually matter stackless python only |
| 385s | does one thing at a time the other |
| 387s | terrifying part is that it's not only on |
| 388s | a single core it's on a single logical |
| 390s | processor |
| 392s | so we're reinforcing fleet fight nodes |
| 394s | for example we're basically just trying |
| 396s | to throw as much clock speed at that |
| 398s | node as possible it doesn't even |
| 400s | actually matter how many cores we put on |
| 401s | it |
| 402s | and that's kind of one of the ultimate |
| 403s | limiting factors that had us start the |
| 405s | conversation about |
| 406s | what became the idea behind quasar and |
| 409s | how we start teasing that problem apart |
| 412s | um but i o completion ports are also |
| 415s | important in the sense that it it it |
| 417s | makes a step towards |
| 419s | deferring the management of those |
| 421s | sockets from python because the more |
| 423s | stuff that we get out of python the more |
| 425s | one thing at a time that we can do |
| 427s | um and so that defers to a kernel and |
| 430s | auto completion force is a is a nifty |
| 432s | trick it's similar to um |
| 435s | polling uh socket polling in linux if |
| 437s | you're familiar with that um there's a |
| 439s | push pull paradigm asymmetry there um |
| 443s | but that's how it works on windows oh by |
| 444s | the way all of this is on windows |
| 446s | um |
| 451s | um and so |
| 453s | as this evolved and and these players |
| 456s | grew and more and more things started |
| 458s | being built inside of this is that part |
| 460s | of like the python that we were talking |
| 461s | about grew and grew and grew under a |
| 462s | nice little folder called script which |
| 465s | basically became a large portion |
| 467s | of of the of the actual code base |
| 470s | what would end up happening is during |
| 471s | releases |
| 473s | um people started noticing that like |
| 475s | forms would go down or web pages would |
| 477s | go down like i think the eve wiki at |
| 479s | that time would would go down |
| 481s | because every release would publish |
| 483s | information about the details of eve so |
| 485s | i think back then we didn't have the sde |
| 487s | and we didn't have any of the apis |
| 489s | and so the reflex to that was to build |
| 491s | the xml api they're like hey this is |
| 494s | this is taking down our web servers and |
| 495s | that's bad because when we do releases |
| 497s | we want everybody to see all the new |
| 498s | information everything we're doing |
| 500s | so they built up the xml apis to kind of |
| 502s | protect from that |
| 503s | and what i |
| 505s | i have i haven't really found the exact |
| 507s | person or the exact reason this came |
| 509s | into existence that's the original |
| 511s | reason for it but i don't think people |
| 513s | realized at that point in time that what |
| 515s | they were doing was effectively making |
| 516s | one of the biggest retention mechanisms |
| 519s | in eve because it allowed you guys to |
| 521s | build on top of that then things like |
| 523s | evmon were born |
| 525s | importing things through like eft and |
| 528s | all those kind of things when we had all |
| 529s | that static data but |
| 531s | ultimately |
| 533s | this was all done xml over http um and |
| 537s | it was it was basically a read cache |
| 539s | nothing nothing fancy there and i think |
| 540s | if anybody remembers the xml api the |
| 542s | cache timers on that were horrendous |
| 545s | it was i think multiple hours on the |
| 547s | actual liveness of the data |
| 551s | and so |
| 553s | this got into things like uh |
| 555s | managing skill plans through through |
| 557s | even |
| 558s | and and you can kind of see an echo of |
| 560s | that with skill plans in the game right |
| 562s | now which oddly enough is connected to |
| 563s | quasar so |
| 565s | it's |
| 566s | serendipitous that one of the first |
| 568s | full features that was 100 on quasar is |
| 572s | actually the same third party developed |
| 575s | feature that was built outside with the |
| 578s | original xml api |
| 581s | and so as this kept growing we kept |
| 584s | adding more and more things to the |
| 585s | ecosystem we started getting |
| 588s | oauth 2 because then we had added a |
| 590s | launcher and we needed other websites to |
| 591s | federate with other information that |
| 592s | might be there |
| 594s | and so you know then we started getting |
| 596s | oauth 2 over http and that ecosystem |
| 598s | started to grow |
| 600s | and i've kind of simplified this so that |
| 601s | monolith services kind of represents |
| 603s | souls and proxies that's the the og |
| 605s | cluster if you will all services started |
| 607s | growing into a suite of dot-net |
| 610s | applications um for managing |
| 614s | uh payment information or various other |
| 616s | things in this suite of uh of services |
| 618s | was actually |
| 620s | technically speaking the first |
| 622s | external service to the original cluster |
| 625s | and so they started dealing with a lot |
| 627s | of the problems without understanding |
| 629s | or not they understood it without really |
| 632s | knowing about kind of the paradigm that |
| 634s | a lot of people talk about today when |
| 635s | people talk about |
| 637s | monoliths to microservices like that is |
| 638s | a thing that almost everyone has heard |
| 640s | something about at this point in time |
| 643s | only a handful of people knew about this |
| 645s | or had completed it right the entire |
| 647s | world was at this point in time dealing |
| 650s | with this problem and nobody had a name |
| 651s | for it yet |
| 653s | and so this was roughly the first |
| 654s | cluster of services that we had live |
| 656s | outside the actual eve ecosystem in |
| 658s | their own sustained way and |
| 661s | also to note all of this stayed inside |
| 663s | of our data center so this was still |
| 664s | inside of |
| 666s | basically metal boxes right next to each |
| 667s | other |
| 669s | but the problem with those is and this |
| 671s | is one of the problems with people that |
| 673s | implement micro services |
| 675s | is that ultimately like yeah we got a |
| 676s | bunch of micro services and they're all |
| 677s | connected to the same database oh that's |
| 679s | not how that's supposed to work |
| 681s | like you've just created another |
| 682s | monolith for your database right |
| 684s | um and this is ultimately the trap they |
| 686s | fell in right |
| 687s | that was one of the tricky things |
| 692s | and so |
| 694s | as we started looking more into kind of |
| 696s | like the surface area of how these |
| 698s | things were growing |
| 700s | this then came on to introducing crest |
| 703s | and my history here is a little blurry |
| 706s | because i was actually coming into ccp |
| 708s | when dust was winding down |
| 710s | um but ultimately crest was built |
| 712s | because of the concept of dust |
| 714s | ultimately the the orbital bombardment |
| 716s | was done through a crest call because |
| 718s | that went from the psn |
| 719s | uh uh network into our network and |
| 722s | that's what basically coordinated the |
| 724s | the orbitable margin strikes um |
| 727s | the interesting thing about crest is |
| 729s | that it was then trying to adopt |
| 731s | mentalities at that point in time and in |
| 733s | this case it was very very academic |
| 735s | mentalities this was a hypermedia |
| 737s | restful json if anyone remembers what |
| 739s | that is it's basically a |
| 740s | self-documenting self-referencing api |
| 744s | and the idea behind that was that a |
| 746s | robot could then navigate the api and do |
| 749s | what it wanted to do |
| 752s | humans unfortunately were doing that not |
| 753s | robots |
| 754s | so that became a lot of paperwork just |
| 757s | to use the api and then on top of that |
| 759s | dealing with all the changes and the |
| 761s | breakages there |
| 762s | the implementation here got more |
| 764s | interesting those |
| 766s | those nodes for crest were effectively a |
| 768s | soul node with more logic on them and so |
| 771s | they were susceptible to the same |
| 773s | scaling problems of basically the one |
| 775s | hop mesh network and also implemented on |
| 778s | top of stackless python |
| 781s | so |
| 782s | it also paved the way for right |
| 784s | endpoints i think crest is the i can't |
| 786s | remember the first right endpoint it |
| 787s | might have been like autopilot |
| 789s | or eve mail i can't remember exactly the |
| 792s | first one |
| 793s | but this also paved the way internally |
| 795s | for this because this was a big deal |
| 798s | not many people cared and i used quotes |
| 800s | there because i'm not saying that nobody |
| 802s | cared about it but they weren't really |
| 803s | bothered by the fact that people were |
| 805s | scraping data and not taking down the |
| 806s | service sounds great when we started |
| 808s | building up crescent like hey we're |
| 809s | gonna allow players to |
| 811s | automatically affect things through the |
| 814s | api |
| 815s | and then it became an entire civil war |
| 817s | internally on like what that meant and |
| 819s | how we should go about doing that |
| 820s | obviously the cat's out of the bag but |
| 822s | at that point in time |
| 824s | it was more about |
| 826s | isolating it to the difference between |
| 828s | what we could affect in the universe and |
| 829s | what was localized to the player |
| 832s | so that still stays to this day there's |
| 834s | not anything you can do even in esi |
| 837s | where you can affect the universe |
| 840s | it has the illusion of that but you |
| 842s | don't actually affect the universe until |
| 843s | we introduce things like |
| 845s | actually manipulating market calls that |
| 847s | affect inventory then things start |
| 849s | affecting the universe but most |
| 850s | everything if i remember correctly |
| 852s | is about endpoints that can only affect |
| 854s | the state of your character like |
| 856s | autopilot contacts eve mail |
| 860s | i can't remember all of them um |
| 863s | and so this set that precedence there |
| 866s | and this kind of introduced yet another |
| 867s | point that we were |
| 869s | building on and so |
| 871s | the problem with this growing surface |
| 873s | area |
| 874s | ultimately became performance like we |
| 876s | were talking about this is all built on |
| 877s | top of stackless python we could only |
| 879s | scale vertically not really horizontally |
| 882s | because the more nodes that we scaled |
| 884s | horizontally the less connections we |
| 886s | could deal with up front and that became |
| 888s | a problem that equation basically didn't |
| 889s | work out i think when we did the math on |
| 892s | that originally that number came to |
| 893s | around a hundred thousand |
| 895s | um |
| 896s | now we haven't been to that number yet |
| 898s | um but uh ultimately that was the the |
| 901s | proponent of that like that was what was |
| 904s | powering those decisions |
| 906s | um |
| 908s | and this gets into cyclist python oh the |
| 910s | gill right so python in general rather |
| 912s | stackless or not doesn't matter has what |
| 914s | they call a global interpreter lock this |
| 916s | is what forces it to do the one thing at |
| 917s | a time but it also makes it very |
| 919s | powerful in the sense that you don't |
| 921s | have the complications of any |
| 923s | concurrency paradigms or primitives that |
| 925s | you then have to coordinate there's no |
| 926s | synchronization ultimately because it's |
| 928s | only doing one thing at a time |
| 932s | the database was also a problem here |
| 935s | um |
| 936s | a big part of why we implemented a lot |
| 938s | of the tools that we have today is |
| 940s | because when we started introducing easy |
| 943s | it basically became the scapegoat for |
| 945s | any problem that came up at any point in |
| 947s | time |
| 948s | to the point where i had a tally board |
| 949s | in the office of not easy or easy |
| 954s | and so the database becomes a bottleneck |
| 955s | for this because ultimately it's the |
| 956s | same problem |
| 958s | we have all of this concurrency |
| 960s | happening at a single location that can |
| 962s | only scale up to a certain degree and |
| 964s | that's that's why you read all the dev |
| 965s | blogs that we have even the recent one |
| 967s | about the hardware upgrades where we |
| 969s | literally have to throw metal at it to |
| 971s | solve some of those problems because the |
| 973s | complexity or the density of the actual |
| 975s | operations being done in the database |
| 976s | can only be mitigated by |
| 979s | faster light in this case |
| 983s | maintenance uh is another big one here |
| 987s | in order to change anything |
| 989s | in xml api that was great because it was |
| 990s | a standalone service we didn't have to |
| 992s | worry about |
| 993s | tranquility going up or down however |
| 996s | it had the side effect of if anyone |
| 998s | changed anything in the database the xml |
| 1001s | api didn't know about it so there was a |
| 1003s | lot of thrashing in the sense that |
| 1005s | endpoints would go down and break and |
| 1007s | various other things would mismatch with |
| 1009s | certain attributes or whatever the case |
| 1010s | may be we still have this problem with |
| 1012s | easy right now in various different |
| 1014s | places that we're still combating but in |
| 1015s | a different way |
| 1017s | um |
| 1018s | this gets into deployments um |
| 1020s | crest could not be affected unless we |
| 1022s | change like brought down tq ultimately |
| 1024s | uh that's one of the other big pieces |
| 1026s | about what we're modernizing it |
| 1028s | um |
| 1029s | uniform criticality this gets back to |
| 1033s | what me and ccp tuxford were talking |
| 1034s | about in vegas and this this talk is |
| 1036s | basically |
| 1037s | a status update of the |
| 1040s | talk we had in vegas |
| 1041s | where we were talking about the concepts |
| 1043s | behind this and more the technology that |
| 1044s | we're using |
| 1045s | and the developer experience that we're |
| 1047s | targeting uh less about where we're at |
| 1049s | now |
| 1050s | and and ultimately what the cultural |
| 1052s | changes need to be to achieve that |
| 1054s | um and uniform criticality in this sense |
| 1056s | means that |
| 1058s | everything is priority one |
| 1060s | and that's a problem |
| 1062s | hey email's not working |
| 1063s | well email could not work to a point |
| 1065s | where it starts cascading failures |
| 1067s | inside the cluster |
| 1069s | well now email is definitely priority |
| 1071s | one but that's the silliest thing to |
| 1073s | have as priority one we would rather |
| 1074s | just turn off email instead and deal |
| 1077s | with that problem and then turn it back |
| 1079s | on when it's ready to go oh |
| 1081s | unfortunately eve's not built in this |
| 1082s | way however our teams have become |
| 1084s | exceedingly efficient |
| 1086s | at |
| 1087s | building and working in this way but |
| 1089s | that is an immense slowdown into how |
| 1091s | they build into what they can |
| 1093s | build and then we get into the |
| 1094s | development aspect of this |
| 1097s | domain boundaries became a huge part of |
| 1099s | what we started talking about because |
| 1100s | ultimately |
| 1102s | when you start building something on top |
| 1103s | of a soul node domain boundaries |
| 1105s | instantly get blurry and if i could |
| 1109s | oversimplify what has happened over the |
| 1111s | last 20 years with the eve code base |
| 1113s | you combine things with |
| 1116s | a dynamic language like python |
| 1119s | maybe some of my little personal biases |
| 1120s | there you have the same you have the |
| 1122s | same database you have a single |
| 1123s | deployment mechanism and what happens |
| 1125s | over time is it doesn't matter how well |
| 1127s | you organize or build the code base |
| 1129s | because that's the thing about eve is |
| 1131s | all the core components are are well |
| 1133s | designed in the sense that they're like |
| 1135s | what people used to call service |
| 1136s | oriented architecture which is now |
| 1138s | microservices we've been doing the same |
| 1140s | thing since the 70s everybody just keeps |
| 1141s | calling it a different thing |
| 1143s | [Laughter] |
| 1146s | we found the old guy |
| 1151s | and so |
| 1152s | ultimately it didn't allow you to |
| 1153s | actually build those boundaries |
| 1155s | everything kind of blurred together and |
| 1157s | it became this thread that you had to |
| 1158s | pull at which caused all these side |
| 1160s | effects |
| 1161s | these side effects that no one did |
| 1163s | intentionally it's just kind of how they |
| 1164s | happen because if you can't isolate the |
| 1166s | domains that you're actually working on |
| 1168s | you can't really take responsibility for |
| 1170s | just that piece |
| 1171s | i mean we can talk about how many |
| 1173s | different types of mission systems that |
| 1174s | we have in eve online |
| 1176s | that's because when you go to look at |
| 1177s | them like i'm going to add this thing |
| 1178s | and you look at it and go nope i'm not |
| 1180s | touching that |
| 1181s | because it's connected to so many other |
| 1183s | things and it's almost always easier to |
| 1186s | build something in a separate corner and |
| 1187s | then some other pieces build on that |
| 1189s | eventually tentacles come out of it and |
| 1190s | everything gets woven together right and |
| 1193s | keep in mind this is over the course of |
| 1194s | 20 years right like this is not |
| 1196s | something somebody went you know what |
| 1197s | i'm going to do i'm going to connect |
| 1198s | every domain into a massive monolith |
| 1200s | that's not what people were doing |
| 1202s | it's much a natural evolution of things |
| 1206s | and then this gets into data ownership |
| 1208s | and this goes back into the database and |
| 1209s | how the database works |
| 1211s | when we have things like hot tables or |
| 1213s | poorly planned queries |
| 1215s | it's because of some other services and |
| 1217s | what they might be querying about that |
| 1218s | information |
| 1220s | but that's kind of broken because they |
| 1221s | shouldn't be sharing that that shouldn't |
| 1222s | be part of the problem |
| 1224s | and data leaking out or being consumed |
| 1226s | by anything else should be a problem |
| 1227s | either like for example |
| 1229s | when we have the |
| 1231s | the off services they were actually |
| 1232s | dipping into the same db and connecting |
| 1234s | character information and user |
| 1235s | information |
| 1236s | well that meant that anybody that wanted |
| 1237s | to do anything crazy with characters |
| 1239s | couldn't because now it affected another |
| 1240s | system that they didn't have any actual |
| 1243s | agency over |
| 1244s | and this is why now you're seeing more |
| 1246s | and more changes come through like with |
| 1248s | what ccp nomad was saying earlier about |
| 1250s | like the skills and what we want to |
| 1252s | change in effect there |
| 1253s | there's we're trying to make it easier |
| 1255s | to define those boundaries so that |
| 1257s | people can make more surgical |
| 1259s | foundational changes instead of just |
| 1261s | kind of adding on and trying to sidestep |
| 1264s | what's already there |
| 1266s | and then we get to the cognitive load |
| 1268s | which is more about how the developer is |
| 1270s | working |
| 1271s | and this is what i mean by |
| 1273s | we've conditioned our engineers |
| 1275s | to keep all of this in their head |
| 1278s | when they work on anything |
| 1280s | and if any of you have worked at the |
| 1281s | different like worked on a project that |
| 1283s | has automated testing versus doesn't |
| 1285s | have automated testing there's a |
| 1287s | radically different mental experience |
| 1289s | and motivating factor if it doesn't have |
| 1291s | automated testing |
| 1292s | i'm not incentivized to make it better |
| 1296s | if it does have automated testing i'm |
| 1297s | more incentivized to make it better and |
| 1299s | even make broader sweeping changes |
| 1301s | and going back to the millions of lines |
| 1303s | of code that we talked about earlier |
| 1305s | there's a lot of missing automated |
| 1306s | testing in that but that's also |
| 1307s | something that we're working on right |
| 1308s | but it's still to create a system or |
| 1310s | connect those systems or |
| 1312s | make them manifest in the |
| 1314s | the experience or or |
| 1317s | i regret saying this word but like |
| 1318s | illusion of gameplay right because |
| 1321s | that's what we're kind of getting at |
| 1322s | right we're we're all living in this |
| 1324s | wonderful fantasy of flying a spaceship |
| 1325s | and those kind of things but to make |
| 1326s | that really connect you have to then |
| 1329s | deal with all of these other pieces when |
| 1332s | it really should just be hey let's |
| 1334s | change the way the spaceship flies |
| 1341s | so |
| 1344s | this kind of led us all of these |
| 1346s | different pieces kind of led us into |
| 1347s | what we were talking about |
| 1348s | with the original idea of quasar we |
| 1350s | didn't we didn't know we were going to |
| 1351s | build quasar by the way this was kind of |
| 1353s | an evolution of how things went |
| 1356s | um |
| 1358s | ultimately the origin was the eve |
| 1359s | swagger interface or the open api |
| 1362s | implementation |
| 1364s | and |
| 1365s | when we started working on that the the |
| 1367s | vehicle for that was the actual mobile |
| 1369s | application |
| 1370s | uh the eve companion app or eve portal |
| 1372s | right |
| 1374s | and when eve portal came out it was |
| 1375s | mostly the fear |
| 1378s | was that number one we had all of these |
| 1379s | new devices that would come online that |
| 1381s | weren't necessarily connected to eve and |
| 1382s | it was much easier to connect to all of |
| 1383s | this so we needed a way to protect the |
| 1384s | cluster which meant we couldn't scale |
| 1387s | horizontally with crest because that |
| 1388s | would heat up resources and it |
| 1390s | definitely wasn't going to be an xml api |
| 1392s | because |
| 1393s | xml um |
| 1395s | and so |
| 1397s | we kind of |
| 1399s | discovered |
| 1401s | a way to sidestep |
| 1403s | decades of technical debt by introducing |
| 1406s | a message bus and that's kind of the |
| 1407s | core piece of where quasar started and |
| 1410s | we didn't know this yet but |
| 1411s | ultimately you know if we look at you |
| 1413s | know going back to talking about eve's |
| 1415s | original design like at the core of it |
| 1417s | how it's designed there's roughly about |
| 1420s | if you if you |
| 1421s | so i'm trying to speak about this in the |
| 1423s | sense of like a restful api but if you |
| 1425s | take the core monolith of eve and try to |
| 1427s | actually dissect what's going on there |
| 1429s | there's roughly about a little over 300 |
| 1431s | services internally to just the python |
| 1433s | code base it's talking to itself |
| 1435s | this is roughly 6 000 endpoints compared |
| 1438s | to the 190 that we have for easy and |
| 1441s | that's just to power everything that you |
| 1443s | see in the actual eve client |
| 1445s | and that's hard to keep track of |
| 1448s | when you don't have anything |
| 1450s | dictating what the domain boundaries are |
| 1452s | what the data ownership is |
| 1454s | and so that was the big reason why we |
| 1455s | chose things like swagger spec which |
| 1457s | eventually became open api because we |
| 1459s | wanted people to be able to have the |
| 1460s | conversation about |
| 1462s | what is it that you actually own what |
| 1463s | are you building against what's your |
| 1465s | contract that you're going to maintain |
| 1466s | for everyone else |
| 1472s | then ultimately we got into |
| 1474s | kubernetes in the cloud space with this |
| 1477s | um |
| 1478s | what ended up happening was |
| 1481s | we were trying to build things against |
| 1483s | our data center against heart like we |
| 1485s | at one point in time we were like pixie |
| 1487s | booting machines into ibm blade centers |
| 1489s | and running cube before v1 |
| 1492s | um it worked it worked but it was not |
| 1496s | sustainable unfortunately |
| 1497s | um and then we kind of just one clicked |
| 1500s | into gke uh inside of google cloud |
| 1503s | that's where we kind of started our |
| 1504s | journey with with kubernetes |
| 1507s | um and that just allowed us to provision |
| 1510s | resources that we would never have |
| 1511s | access to |
| 1512s | to wield a lot of power that we would |
| 1514s | never have access to with |
| 1516s | things like |
| 1517s | we wouldn't have to worry about the link |
| 1519s | speed |
| 1520s | of what's coming into our data center uh |
| 1522s | versus just making a load balancer and |
| 1524s | everything coming in we eventually |
| 1525s | landed on amazon for various other |
| 1528s | reasons but |
| 1529s | that's kind of the journey that took us |
| 1531s | there |
| 1532s | and then ultimately the message bus was |
| 1533s | the core piece of this |
| 1535s | and |
| 1536s | we chose the message bus over a service |
| 1539s | service mesh architecture because the |
| 1541s | ideas that we had about how this would |
| 1543s | evolve |
| 1546s | up front a service mesh is very |
| 1549s | difficult to get all the right tooling |
| 1551s | in place to help people debug and |
| 1553s | maintain whereas a message bus gives you |
| 1555s | a bottleneck which seems |
| 1557s | counterintuitive in the grand scheme of |
| 1558s | things but gives you a dedicated |
| 1560s | bottleneck to own all the pieces that |
| 1562s | are flowing through there and allows |
| 1563s | your like your upfront cost as far as |
| 1565s | getting other teams on board |
| 1567s | to go faster sooner um so this is the |
| 1571s | distinction that we made originally this |
| 1572s | is also while the world was still |
| 1573s | figuring out things like istio |
| 1575s | linker d |
| 1577s | envoy ambassador all the other cool |
| 1579s | things that are out there now |
| 1582s | and we still we still talk about this |
| 1583s | heavily because we're now to the point |
| 1585s | where we're emulating a service mesh to |
| 1587s | a degree |
| 1588s | but we ultimately wanted the teams to |
| 1590s | not have to worry about what the ingress |
| 1592s | looked like we only wanted the teams to |
| 1594s | worry about their domain |
| 1595s | and the data they owned so how do we |
| 1598s | make it so that they only care about |
| 1599s | inputs and outputs that was our primary |
| 1601s | goal |
| 1603s | this led us to protobuf |
| 1605s | after |
| 1606s | doing everything in eve portal and with |
| 1609s | esi and there's still esi endpoints that |
| 1610s | do things through json sorry let me be |
| 1612s | clear |
| 1613s | all of you guys see json on the back end |
| 1616s | we see some endpoints that are doing |
| 1618s | protobuf and we see some endpoints that |
| 1619s | are doing json |
| 1621s | when we started to build and just |
| 1622s | basically blitz through the easy spec |
| 1625s | and started building everything we built |
| 1626s | it all in json |
| 1628s | we learned real fast that was going to |
| 1630s | be a problem when we didn't have a |
| 1631s | schema to really deal with wrangling in |
| 1633s | all the data and that's kind of why we |
| 1635s | started looking at things like protobuf |
| 1637s | and we started looking at things like |
| 1638s | protobuf to deal with |
| 1640s | performance as well once we realized oh |
| 1642s | protobuf has this nice |
| 1643s | uh c-plus plus uh mechanism where you |
| 1646s | can generate native code that can also |
| 1648s | do the serialization for protobuf and |
| 1650s | what that means is |
| 1651s | we're basically moving everything from |
| 1654s | our do one thing at a time stackless |
| 1656s | python of writing down messages |
| 1658s | and then just throwing that memory at c |
| 1660s | plus plus and saying you do this instead |
| 1663s | while python can go do the next single |
| 1664s | thing it can do which is a huge |
| 1666s | performance benefit for us |
| 1669s | naturally |
| 1670s | this led us to grpc |
| 1673s | because ultimately when we started doing |
| 1674s | this we started connecting to this as a |
| 1676s | server everything was going great this |
| 1678s | is how we established like |
| 1679s | a lot of what you're seeing from the |
| 1681s | data teams a lot of the newer pipeline |
| 1683s | around the definitions of those events |
| 1685s | what's being basically fire hosed out of |
| 1687s | the system is coming through |
| 1689s | protobuf into the message bus ecosystem |
| 1692s | but ultimately when we started talking |
| 1693s | about this more and more we realized uh |
| 1695s | we need a way to actually connect |
| 1696s | between these systems what makes sense |
| 1698s | there we didn't want to maintain a |
| 1699s | protocol for this there was no point in |
| 1701s | that there were so many to pick from |
| 1703s | and protobuf became kind of the the |
| 1705s | anchor for this because it was just a |
| 1707s | hop skip and a jump away and we could |
| 1708s | generate grpc endpoints |
| 1710s | um |
| 1712s | then we realized oh we can put this in |
| 1713s | the client |
| 1714s | and that's where the idea of quasar |
| 1716s | started when we realized wait we can |
| 1718s | close the loop on the entire ecosystem |
| 1720s | and sidestep the entire legacy code base |
| 1724s | and keep everything inside of cube |
| 1726s | inside of golang instead of the message |
| 1727s | bus |
| 1728s | and not have to deal with anything |
| 1730s | that's going i mean we do have to deal |
| 1731s | with it all the time like |
| 1732s | it's not all sunshine and rainbows we |
| 1734s | still have to go in and make sure things |
| 1736s | connect and actually manifest in the |
| 1738s | universe |
| 1739s | the way they're supposed to |
| 1741s | and this then got us into domain |
| 1743s | services |
| 1744s | and |
| 1745s | i'm i i really don't like the word micro |
| 1748s | services number one because no one knows |
| 1750s | what it means uh but also number two it |
| 1752s | defines an arbitrary scope to what |
| 1755s | you're designing |
| 1757s | and this is why we talk about domain |
| 1758s | services because it ties it back to the |
| 1760s | actual data model like what are you |
| 1762s | actually building and what should you |
| 1764s | own |
| 1765s | an example of this and one of the first |
| 1767s | kind of domain servers that we built for |
| 1768s | eve was skill plans |
| 1770s | skill plans owns everything that it does |
| 1772s | and it never touches the monolith at all |
| 1775s | other than sending out by by proxy of |
| 1778s | sending out like other events of like |
| 1779s | hey they want to train this skill now |
| 1781s | from from the skill plans |
| 1783s | um even that might be debatable it might |
| 1785s | be going through the client point is all |
| 1787s | of that data that you're sharing with |
| 1789s | your corporation with all those skill |
| 1790s | plans all those different pieces |
| 1792s | that's all completely going through |
| 1794s | quasar |
| 1795s | and we have some other services before |
| 1796s | that where they're going through quasar |
| 1798s | the activity tracker is another one of |
| 1799s | them but it wasn't quite doing the same |
| 1801s | thing |
| 1802s | and we can kind of point that out here |
| 1805s | um |
| 1806s | oh wait a minute |
| 1809s | yes |
| 1811s | um |
| 1813s | so this is kind of where we're at now |
| 1815s | those are tiny words um |
| 1819s | yeah |
| 1820s | and so this kind of represents where |
| 1821s | quasar is in in kind of the cloud |
| 1823s | provider that we have uh we ultimately |
| 1825s | have a service gateway which is the |
| 1827s | first piece of this puzzle and that is |
| 1829s | our authoritative domains this is like |
| 1831s | if there's an event inside of this |
| 1833s | uh domain it is a fact of the universe a |
| 1836s | ship exploded uh this guy bought |
| 1838s | something on the market whatever the |
| 1840s | case may be and this is what's normally |
| 1841s | referred to as east-west traffic in the |
| 1844s | terms of kind of your network topology |
| 1846s | this is usually within owned |
| 1849s | networks uh for that for that company |
| 1851s | um |
| 1853s | and you can kind of see here where we |
| 1854s | introduced the mobile client all these |
| 1855s | other pieces that |
| 1856s | eventually got pieces of quasar it |
| 1858s | wasn't known as quasar at that point in |
| 1859s | time but eventually got in |
| 1861s | the public gateway then represents our |
| 1862s | north-south traffic uh which is |
| 1864s | basically anything that egresses or |
| 1866s | ingresses between controlled networks so |
| 1869s | basically your guys's machine versus our |
| 1871s | guys's machine |
| 1873s | and those we treat radically differently |
| 1875s | because if we |
| 1876s | emit an event on the service gateway |
| 1878s | it's a fact if your client emits an |
| 1881s | event it needs to be statistically |
| 1882s | significant |
| 1884s | and what i mean by this like when we're |
| 1886s | tracking like how people use certain |
| 1888s | things within the client like opening or |
| 1890s | closing windows or the case may be we |
| 1892s | can't trust any of that data it's coming |
| 1894s | from an untrusted source and you know |
| 1898s | clients get modified every now and then |
| 1900s | it seems so we have to take into account |
| 1902s | like what is true and what is not so |
| 1903s | they have to be statistically |
| 1904s | significant |
| 1907s | and this is the part where we started |
| 1909s | talking about internally |
| 1911s | where ultimately the desktop client |
| 1913s | isn't the only client and this started |
| 1916s | opening the door for how we talk about |
| 1918s | the future of |
| 1919s | of how eve works and what happens and |
| 1921s | then what we build |
| 1923s | um where we started talking about eve |
| 1924s | portal the websites the third-party apps |
| 1927s | that you guys are constantly building um |
| 1929s | all of those pieces it it means that you |
| 1931s | could play potentially play eve |
| 1934s | from more than just the desktop client |
| 1937s | so part of this was proprietary to |
| 1939s | standards we're talking about things |
| 1940s | like our original like carbon io that |
| 1943s | proprietary python |
| 1946s | protocol going into things like protobuf |
| 1948s | grpc those kind of things over amqp or |
| 1950s | google pub sub or nats or whatever the |
| 1952s | case may be i skipped ahead that's the |
| 1954s | message bus one |
| 1956s | but that ecosystem is also the big part |
| 1958s | of this when we talk about things like |
| 1960s | what ccp no man was talking about with |
| 1962s | the air career program right |
| 1965s | all of those pieces that we're doing |
| 1966s | there aren't specifically for the air |
| 1968s | career program right now they are |
| 1971s | but all of those extra events all of |
| 1972s | those things that are being tracked |
| 1973s | those are all pieces that we can reuse |
| 1975s | within that ecosystem so the more and |
| 1977s | more pieces that we have this is kind of |
| 1978s | the original uh |
| 1980s | ignition of of the activity tracker |
| 1982s | while the activity tracker didn't act on |
| 1984s | these things it still tracked all of |
| 1985s | them and then it gave us all of this |
| 1987s | extra information on how to react and |
| 1989s | how to build upon that information |
| 1991s | that's already been throwing around the |
| 1993s | the uh the ecosystem |
| 1995s | and and we prototyped these a while back |
| 1996s | i think |
| 1997s | i think there was even a fan fest where |
| 1999s | we put up arbitrary data or that |
| 2001s | arbitrary data we put a prototype data |
| 2003s | on a kill mail system and everybody lost |
| 2005s | their minds over logic or logic info on |
| 2007s | the uh on the kill mails um |
| 2010s | that is also something that we're |
| 2011s | looking to do and proceed to but like |
| 2013s | that's part of this evolution and part |
| 2014s | of the performance pieces that we're |
| 2016s | talking about here |
| 2018s | this also gave us a ubiquitous language |
| 2020s | this was one of the biggest problems |
| 2022s | that we had internally |
| 2023s | you could build a service or any of |
| 2025s | those pieces and go to a separate team |
| 2027s | and then go look at that and go i can't |
| 2028s | use that when they really could but |
| 2030s | there was no ubiquitous language to |
| 2031s | communicate that so protobuf gives us |
| 2034s | that ubiquitous language in the entire |
| 2035s | ecosystem where we can go and say hey |
| 2037s | i'm going to make a call here and that |
| 2039s | service doesn't care who it is or what |
| 2041s | it's for it doesn't have to care about |
| 2043s | something inside of that python module |
| 2045s | mutating it to something else or |
| 2047s | changing something that shouldn't need |
| 2048s | to or somebody else deploys a different |
| 2049s | version of that our teams are now |
| 2051s | building around the concept of you own |
| 2053s | this api you need to keep this api |
| 2055s | working and if we want to change that |
| 2057s | that's a conversation around the actual |
| 2059s | language and the domain which then gets |
| 2061s | us to our domain services |
| 2063s | which is the piece more around |
| 2065s | what do you own |
| 2067s | what do you iterate on those kind of |
| 2068s | things an example of this is skill plans |
| 2070s | again |
| 2071s | where |
| 2073s | we were talking about modifying how |
| 2076s | skills work where it's not no longer a |
| 2078s | cue you're dumping skill points into |
| 2080s | uh like you're accruing skill points and |
| 2083s | then you do with those what you want you |
| 2084s | don't have to actually plan that out |
| 2086s | and the evolution of skill plans might |
| 2088s | be |
| 2089s | that it just becomes the domain service |
| 2090s | for skills that might be the natural |
| 2092s | evolution of that |
| 2094s | we've yet to see that because we're |
| 2095s | still learning these pieces and again |
| 2097s | these are the services that are still |
| 2099s | kind of the first ones of their kind |
| 2101s | inside of quasar |
| 2104s | so ultimately kind of what did we learn |
| 2105s | from this |
| 2106s | it gets into the micro versus domain |
| 2108s | um |
| 2109s | and this kind of gets into |
| 2111s | the delineation be |
| 2113s | because like |
| 2115s | the the biggest problem like if you |
| 2117s | think about it |
| 2118s | abstractly when you have a game engine |
| 2120s | involved in anything |
| 2122s | that is instantly a monolith the client |
| 2123s | for that for that game is a monolith |
| 2125s | there's not much you can do around that |
| 2127s | there's a lot of talks around that you |
| 2129s | hear about micro uis or micro front ends |
| 2132s | or those kind of things that might be |
| 2133s | the next evolution that we'll see |
| 2136s | but this is basically the difference |
| 2137s | between what we like we couldn't use any |
| 2139s | of these technologies in eve because |
| 2141s | all of those things were detached like |
| 2142s | if you use spotify the little bar at the |
| 2144s | bottom of it was its own |
| 2146s | http call that went to a separate |
| 2147s | service whereas an eve that connects to |
| 2150s | the proxy which goes with the soul nodes |
| 2151s | routes over information goes to the same |
| 2152s | database and comes back through |
| 2154s | everything was connected right and so |
| 2156s | that's the big difference for us and we |
| 2157s | want to concentrate on the domains uh |
| 2159s | not the individual mechanisms |
| 2162s | and then learning the difference between |
| 2164s | a message bus and a service mesh |
| 2166s | kind of getting to the nuances of |
| 2168s | dealing with connectivity ingress how |
| 2170s | players connected and kind of getting |
| 2173s | that off the table so our devs could |
| 2175s | concentrate on other things |
| 2176s | and then getting to api |
| 2178s | representing the team boundary |
| 2180s | not like i you know we kind of have |
| 2182s | evolved from this uh building features |
| 2185s | in the sense of i need to build all of |
| 2187s | these pieces because i need all the |
| 2188s | pieces of this feature in order to make |
| 2189s | this thing which the side effect over |
| 2191s | that over time is you have a lot of |
| 2193s | things that are very similar and you |
| 2195s | don't evolve the existing ones |
| 2197s | as opposed to we own the api for |
| 2199s | characters do you need more data need to |
| 2201s | change the way that something works then |
| 2203s | there's a team that can have a |
| 2204s | conversation with that and usually |
| 2205s | that's over a pr over protobuf |
| 2208s | and ultimately |
| 2210s | new technology is easy culture is not |
| 2212s | and i and i say that it's a relative |
| 2214s | statement like there's a lot of complex |
| 2216s | things that we're doing with technology |
| 2218s | but the thing that surprised us the most |
| 2220s | was kind of people's reaction to that |
| 2221s | new technology |
| 2223s | some people jumped right in other people |
| 2224s | it kind of reflected some deficiencies |
| 2226s | that we had and kind of the processes |
| 2228s | that we were doing were again going back |
| 2230s | to automated testing where we were |
| 2231s | pulling people into the spotlight of |
| 2233s | like cool where's your test and i'm |
| 2234s | going i don't have any you can't you |
| 2236s | have to add those things in this |
| 2238s | ecosystem and so evolving that culture |
| 2241s | to understand like what the progress of |
| 2242s | those types of things would be |
| 2245s | and this is the question i've gotten on |
| 2247s | different podcasts and streams that i've |
| 2248s | talked on |
| 2250s | why why concentrate on this why not |
| 2252s | build more features when i do all of |
| 2253s | this like |
| 2255s | this is a holistic approach to how we |
| 2258s | need to fundamentally fix a lot of |
| 2260s | different things in the ecosystem over |
| 2262s | time over 20 years of teams isolating |
| 2265s | and what the features that they only |
| 2266s | need to build |
| 2267s | and kind of the the turbulence and |
| 2269s | natural ups and downs of a company and |
| 2272s | people and people's lives in real life |
| 2274s | and those kind of things |
| 2275s | ultimately we need to fundamentally |
| 2277s | change |
| 2278s | how we're working and in order to do |
| 2280s | that we need to change the technology of |
| 2282s | what we're actually building upon |
| 2284s | because if we need to fundamentally we |
| 2286s | need to fundamentally change how eve |
| 2287s | works and we can't do that unless we |
| 2289s | change how we work |
| 2291s | um and so quasar is kind of the |
| 2292s | fundamental stepping stone uh that we're |
| 2295s | using to build more and maintain more of |
| 2297s | the the eve universe |
| 2300s | the end |
| 2301s | thanks |
| 2302s | [Applause] |
| 2313s | do we do questions here |
| 2317s | ah |
| 2319s | i couldn't tell if that was somebody who |
| 2320s | had the authority to say that or not |
| 2326s | yeah go ahead the old |
| 2329s | guys daddy |
| 2331s | [Laughter] |
| 2340s | ah |
| 2341s | yeah |
| 2343s | yeah so the question is like nadia the |
| 2344s | new graphical editor |
| 2346s | uh that's using to build a lot of the |
| 2348s | content is it using quasar it is not |
| 2350s | specifically using quasar because the |
| 2351s | majority of that is client-side |
| 2353s | experience mechanisms that's going on |
| 2355s | but it is hooked into a lot of the event |
| 2357s | loops that are flying into quasar so we |
| 2360s | can observe a lot of what's happening |
| 2362s | there and so as that team does more and |
| 2364s | more that's kind of outside a |
| 2366s | unique uh experience for a single player |
| 2370s | because that's ultimately like so far |
| 2372s | the np is that um once it starts it |
| 2375s | kind of going outside of that scope it |
| 2377s | will probably wander more into our |
| 2379s | territory as far as what we need to |
| 2380s | support |
| 2384s | you talked about uh |
| 2396s | i wouldn't say blameless |
| 2398s | no so i mean we so we try to do we try |
| 2401s | to do retros for that and i would argue |
| 2403s | that a lot of the the team that works on |
| 2404s | quasar and the infrastructure teams in |
| 2406s | general um |
| 2408s | there are |
| 2409s | elements of sre there |
| 2412s | where |
| 2413s | so we do we do rotations on call |
| 2415s | rotations but we kind of combine that |
| 2417s | with like if you're on call |
| 2419s | i'm not gonna care if you don't get your |
| 2421s | primary project done i want you to |
| 2422s | concentrate on like answering people's |
| 2424s | questions of course if there's alerts |
| 2425s | something melts down |
| 2426s | all those kind of things but if |
| 2428s | everything is quiet it's kind of one of |
| 2430s | those things of what's making the most |
| 2431s | noise make it stop making noise |
| 2433s | so we kind of have that sre mentality in |
| 2435s | that sense |
| 2437s | the other aspect of that is we're big |
| 2438s | fans of slos |
| 2440s | um and trying to keep track of those |
| 2442s | things and seeing things before they |
| 2444s | catch fire like being able to see the |
| 2445s | smoke before the fire is pretty powerful |
| 2448s | and that has to do with a lot of the |
| 2449s | tooling that we've introduced the |
| 2450s | ecosystem not only just quasar but the |
| 2452s | the original code base of people in line |
| 2453s | with things like |
| 2455s | sentry honeycomb grafana prometheus |
| 2458s | there's tons of stuff like that yeah |
| 2461s | so the |
| 2463s | simulation obviously still needs to run |
| 2471s | no thankfully not |
| 2473s | i checked |
| 2475s | so |
| 2477s | what like what |
| 2479s | is plan in terms of |
| 2481s | splitting out more and more of those |
| 2483s | non-simulation |
| 2485s | services into those domain services |
| 2487s | right so the question is is what is kind |
| 2489s | of the forward plan of simulation |
| 2493s | based services versus |
| 2495s | non-simulation-based services ultimately |
| 2498s | this was the original idea |
| 2501s | people don't necessarily agree with me |
| 2502s | on this but i don't |
| 2504s | usually call eve complex it's very dense |
| 2507s | there's just a lot there |
| 2508s | like if you look at email it's not |
| 2510s | complex |
| 2511s | but in the grand scheme of things it's |
| 2513s | it there's just a lot more going on what |
| 2515s | it can interact with and those types of |
| 2516s | things |
| 2517s | so the general idea was that we clear |
| 2519s | the table |
| 2520s | of all of these services so that people |
| 2522s | could think a bit bigger about what they |
| 2524s | could actually build and even that is |
| 2525s | kind of the core of what our team is for |
| 2527s | we're we're supposed to be a force |
| 2528s | multiplier for the developers and the |
| 2531s | more specifically the feature teams that |
| 2532s | are building all the stuff that you guys |
| 2534s | actually do on a daily basis like this |
| 2536s | is why i say cosmic plumber if you know |
| 2538s | that it's a problem then you don't ever |
| 2540s | think about the plumbing in your house |
| 2541s | until it's busted right |
| 2542s | um |
| 2544s | with the simulation pieces |
| 2546s | we have some theories about that that |
| 2547s | we're very interested in testing |
| 2549s | because when we talk about the |
| 2550s | performance piece of quasar that's kind |
| 2552s | of one of the one of the big pieces |
| 2553s | about grpc |
| 2555s | when we look at fleet fights |
| 2556s | like we're estimating around 30 percent |
| 2559s | of the performance there is spent on |
| 2561s | multiplexing serialization and |
| 2563s | transmission well iocp because it defers |
| 2566s | it to the kernel gets rid of a bit of |
| 2568s | that but it doesn't because it still |
| 2569s | needs to interact with the socket |
| 2571s | serialization is still in python which |
| 2573s | is very slow |
| 2575s | and multiplexing meaning |
| 2577s | 7000 people in a system one person goes |
| 2579s | bang and there's 6999 other messages |
| 2582s | that we need to send |
| 2583s | what we've done with quasar with grpc |
| 2586s | mechanisms in the server is that it's |
| 2588s | offloaded to a separate thread so |
| 2589s | basically because of the python to |
| 2593s | c-plus plus mechanism in protobuf that's |
| 2595s | just comes stock with it we just have to |
| 2597s | marshal |
| 2598s | memory over and then we have a separate |
| 2600s | thread |
| 2601s | crazy eve has a separate thread to do |
| 2604s | something where it's actually doing the |
| 2606s | serialization of the transmission so we |
| 2608s | get it for free |
| 2611s | and then we lean heavily on the message |
| 2613s | bus ecosystem which is where the dynamic |
| 2616s | mechanisms come in and there's a bigger |
| 2617s | conversation that we could have around |
| 2618s | like |
| 2619s | if you'd be surprised the the features |
| 2622s | that require the most complex routing |
| 2624s | mechanisms one of them that highlights |
| 2626s | this is shared bookmarks |
| 2628s | from a routing perspective that becomes |
| 2629s | a nightmare real fast |
| 2631s | and it's one of the few things that are |
| 2633s | actually implemented on the proxy side |
| 2634s | because it needs all of that information |
| 2637s | i don't answer your question |
| 2641s | i think more of it |
| 2645s | is that like would it make sense for a |
| 2647s | team to say you know what i want to get |
| 2648s | |
| 2653s | right |
| 2654s | so would we would we preemptively move |
| 2656s | things over into quasar um yes if |
| 2659s | there's a vehicle for it we're just not |
| 2662s | there to go and like we're gonna |
| 2663s | refactor everything |
| 2665s | no one's gonna sign up for that right um |
| 2669s | so it comes with a vehicle like what are |
| 2671s | we doing like for example a lot of the |
| 2673s | work that we've done under the hood for |
| 2675s | chat to get it out of xmpp in the |
| 2677s | current state that it's in |
| 2679s | is that behind the scenes we've had to |
| 2680s | build a present service |
| 2683s | which has to know where people are in |
| 2685s | eve at all times authoritatively which |
| 2688s | hilariously we found is very difficult |
| 2691s | um so but we need something to motivate |
| 2693s | those types of changes then we'll go |
| 2694s | back into them like if if the if the |
| 2696s | skills revamp starts touching on things |
| 2698s | like characters and there's enough |
| 2700s | traffic there we might want to pull |
| 2701s | characters into a service instead of |
| 2702s | quasar but that would be |
| 2705s | significant open-heart surgery right |
| 2707s | so what you're saying is |
| 2710s | our services |
| 2714s | yes go that way |
| 2719s | uh i saw a talk a few years ago |
| 2722s | similarly titled |
| 2723s | and uh |
| 2724s | the speaker talked about adopting |
| 2727s | my memory |
| 2730s | yeah was it did it have an orange bus |
| 2732s | icon in this in the |
| 2734s | yeah that was being tuxford in vegas |
| 2749s | oh i could ramble about this for a while |
| 2750s | but the short version of that is |
| 2751s | basically beam is cube before cube was a |
| 2754s | thing |
| 2755s | um and the only difference really is |
| 2757s | kind of the api |
| 2759s | and this is kind of the trend that i'm |
| 2761s | seeing in technology in general is that |
| 2763s | the implementation doesn't really matter |
| 2764s | it's the apis that matter so prometheus |
| 2766s | for example everyone loves the apis for |
| 2768s | prometheus and how to aggregate data and |
| 2770s | how to transmit remote data those kind |
| 2772s | of things everyone also hates the |
| 2774s | implementation of prometheus because it |
| 2776s | eats all the ram and most people don't |
| 2778s | take into consideration cardinality and |
| 2780s | those types of things so for things like |
| 2781s | erlang elixir and beam like that whole |
| 2783s | ecosystem is actually quite amazing but |
| 2785s | it's not compatible with anything |
| 2788s | current in that sense and it's also |
| 2790s | doesn't provide a good |
| 2792s | external control plane with kind of what |
| 2794s | the rest of the world is used to i think |
| 2795s | that's the big difference like we tried |
| 2797s | running you know |
| 2799s | uh beam inside of cube but doesn't make |
| 2800s | sense because beam wants to own the |
| 2802s | hardware and then it clusters itself and |
| 2804s | all those nice things but that it |
| 2805s | requires your entire ecosystem to be |
| 2807s | inside of erlang or elixir and that's |
| 2809s | kind of where the success of cube came |
| 2810s | from because it gave it gave everyone |
| 2812s | primitives to to have that ubiquitous |
| 2814s | language to have a conversation across |
| 2816s | basically the entire globe that's why it |
| 2818s | caught |
| 2829s | fire one system and i know there's been |
| 2832s | significant hardware upgrades |
| 2836s | so much as the technology needs to |
| 2838s | develop what have you guys done or what |
| 2841s | you guys think needs to be done to |
| 2849s | so how do |
| 2850s | i'm oversimplifying but how do we make |
| 2852s | things go faster with quasar |
| 2854s | um |
| 2854s | ultimately this kind of goes back to |
| 2856s | we're talking about earlier when we were |
| 2857s | talking about the the effects of |
| 2860s | transmission serialization multiplexing |
| 2861s | those types of things um |
| 2864s | what i'm trying to do and we're toying |
| 2866s | with and then playing with the idea of |
| 2868s | is |
| 2869s | sending simulation frames over quasar |
| 2871s | because we know one that's already |
| 2873s | significantly faster just over the wire |
| 2876s | it's significantly faster |
| 2879s | theoretically we know we can then free |
| 2881s | up 30 of the processing time during a |
| 2884s | massive fleet fight that that is our |
| 2885s | upper bounds of what we could |
| 2886s | potentially bring to the table but that |
| 2888s | is a |
| 2889s | non-trivial project |
| 2891s | and literally reassembling the train as |
| 2893s | it's going down the tracks um so |
| 2896s | we haven't engaged in any of this yet |
| 2898s | and again this comes with the clearing |
| 2899s | the table concept of like we'll keep |
| 2901s | moving things off the table which |
| 2903s | in effect will give us some certain |
| 2905s | percentage of there's less things that |
| 2906s | this needs to do |
| 2908s | but in the grand scheme of things you at |
| 2909s | the end of the day you still wind up |
| 2911s | with a node dedicated to jeta |
| 2913s | and and it doesn't matter how many |
| 2915s | services we take away at that point |
| 2916s | there's still a node dedicated to jita |
| 2918s | even if it's only for the uh simulation |
| 2921s | aspect of it |
| 2923s | um so that's kind of why we're toying |
| 2924s | with the ideas behind like we could send |
| 2926s | simulation frames over this and get a 30 |
| 2929s | bump in in how we're doing things and |
| 2931s | there might be some other things in |
| 2932s | there like the things that we've talked |
| 2933s | about in the past and this is all |
| 2934s | theoretical |
| 2936s | eve is 100 accurate |
| 2938s | but it doesn't necessarily need to be |
| 2939s | and i know that's a terrifying statement |
| 2941s | um |
| 2942s | because when you have seven thousand |
| 2944s | people shooting that one guy at some |
| 2945s | point time you gotta go he's dead like |
| 2947s | stop |
| 2948s | stop counting the bullets he's gone |
| 2951s | uh but eve keeps going yep still dead |
| 2954s | still still dead |
| 2955s | um so there's there's maybe some other |
| 2957s | like philosophical things that we could |
| 2958s | take on like how we deal with the rules |
| 2960s | engine and the simulation in that regard |
| 2962s | but this is all theory crafting because |
| 2964s | again it comes back to |
| 2966s | the the vehicle that we have to move |
| 2967s | forward with those things but i'm i am |
| 2970s | personally chomping at the beds to find |
| 2971s | something to hook that two that we can |
| 2973s | toy with that idea |
| 2974s | it might the first iteration of that |
| 2976s | might be something like we don't send |
| 2979s | like the data that comes in for the gate |
| 2980s | holograms |
| 2981s | like the state on the other side of the |
| 2983s | gate |
| 2984s | we might toy with the idea of routing |
| 2986s | that over quasar like start simple there |
| 2988s | instead kind of again the clear the |
| 2990s | table mentality of like how far deeply |
| 2992s | can we go into that that space |
| 3006s | right so what domain services have been |
| 3008s | put into quasar um |
| 3009s | so |
| 3010s | there is a chat service that we haven't |
| 3011s | rolled out yet it's kind of been a |
| 3013s | shadow service |
| 3014s | some players have already found that |
| 3017s | um |
| 3018s | and skill plans is another one activity |
| 3021s | tracker was the original one but |
| 3022s | activity tracker is not necessarily a |
| 3024s | 100 quasar service in the sense that |
| 3026s | uh it smuggles data through the original |
| 3029s | carbon io connections uh because it was |
| 3031s | built before we had the connectivity to |
| 3033s | the client |
| 3034s | so we're like oh we can consume and |
| 3035s | track all these events that are coming |
| 3036s | in but we can't tell anybody so we just |
| 3039s | sent it back through the server itself |
| 3040s | down to the client that's something that |
| 3041s | we could uh probably renovate or that |
| 3043s | will come with the changes that we're |
| 3045s | doing for the air career program um the |
| 3047s | air career program will be another one |
| 3049s | that's 100 uh quasar uh the one we're |
| 3051s | talking about earlier which isn't really |
| 3052s | player-facing but the |
| 3054s | um |
| 3056s | presence management which we normally |
| 3058s | we're doing in xmpp |
| 3060s | which |
| 3062s | fun fact |
| 3063s | 90 of the traffic in xmpp for us is |
| 3065s | presence not chat |
| 3067s | it's just telling everyone where they |
| 3069s | are that's that's the biggest |
| 3070s | multiplexing problem that we have |
| 3072s | um there's probably some other ones i'm |
| 3074s | forgetting but those are the |
| 3076s | i think uh |
| 3079s | data |
| 3081s | oh yeah for like the data pipelines for |
| 3084s | uh for data and analytics um |
| 3086s | do we still do the recommendation stuff |
| 3089s | yeah |
| 3090s | the recommendation like so the |
| 3091s | recommendations that you get the three |
| 3093s | recommendations that you get if you get |
| 3094s | into is that still feature flagged i |
| 3096s | can't remember |
| 3098s | it's for everybody so yeah so you those |
| 3099s | three recommendations that come in |
| 3100s | that's actually closing the loop from |
| 3102s | the client to quasar to |
| 3105s | the data cube if you will warehouse lake |
| 3108s | i don't know any of the data terms um |
| 3110s | and then that's coming back through |
| 3111s | quasar the client of saying hey you want |
| 3112s | to do one of these three things based on |
| 3114s | what you've been doing in eve |
| 3116s | uh i think that's it you know my guys in |
| 3119s | here that i can |
| 3120s | yeah we'll just stop there |
| 3122s | last question |
| 3124s | that guys |
| 3149s | right |
| 3150s | so the question is like if we have all |
| 3152s | these things emitting events and parts |
| 3153s | of them come down and go back up how do |
| 3155s | we deal with integrity in that regard so |
| 3158s | there are massive papers that you can |
| 3160s | read on that that are really boring but |
| 3161s | event sourcing is the answer to that |
| 3163s | question |
| 3164s | um |
| 3165s | a lot of how we deal with that is mostly |
| 3168s | a little bit more than best effort |
| 3170s | delivery and what i mean by that is best |
| 3172s | effort is usually like it's on the |
| 3173s | socket good luck um |
| 3175s | so we also have a little bit more than |
| 3177s | that where we do a lot of disk queuing |
| 3178s | and mechanisms for like publishing |
| 3180s | confirms with rabbitmq so we basically |
| 3182s | say hey rabbitmq send this to people and |
| 3184s | tell me when the first guy got it and if |
| 3187s | that doesn't happen it goes to disk and |
| 3188s | we retry so this usually manifests in |
| 3191s | that fail state as as a thundering herd |
| 3193s | or a stampede basically um where i think |
| 3197s | we've talked about this on twitter |
| 3198s | during certain interesting situations |
| 3200s | where it's like yep we're now draining |
| 3201s | 50 million events because something fell |
| 3204s | over |
| 3205s | but that's the big difference between |
| 3206s | like i was saying earlier with uh the |
| 3208s | the events in the universe that are |
| 3210s | facts |
| 3211s | those are the ones that we treat uh uh |
| 3214s | with more respect i guess if you will uh |
| 3216s | those are the ones that we trust if that |
| 3218s | thing comes through it's true whereas |
| 3219s | the events that are coming from the |
| 3220s | client you have to be statistically |
| 3222s | significant because if it falls over we |
| 3224s | don't care |
| 3227s | they do that's true yes yes they do yeah |
| 3234s | indeed slo is service level objective uh |
| 3237s | there's also slis which are indicators |
| 3240s | and these are different from alerts in |
| 3241s | general this gets into sre stuff but |
| 3244s | ultimately it's more about |
| 3246s | knowing that your system is trending |
| 3247s | poorly versus something terrible has |
| 3249s | already happened |
| 3251s | yeah i think you said that was the last |
| 3252s | one all right thanks guys appreciate it |
| 3254s | we also have a roundtable after this |