Bestand wählen
Merken

It's Time for Us to Move: The Story and Migrating Hosted Chef to AWS

Zitierlink des Filmsegments
Embed Code

Automatisierte Medienanalyse

Beta
Erkannte Entitäten
Sprachtranskript
the 1st of his in remark Harrison he's a senior sysadmin chef but he is the sysadmin in a company full of CIS admins and you can talk about migrating hosted shift to interview US thank you calendario can you OK I could your hands how many people here use host OK epic Kiki handle a key the old a few new we migrated flashier before this talk if you knew we migrated hosted shaft West last year so what do you mean and if curious how many button you and OK for those who don't know and hosted share phase shift software-as-a-service product graphs and it's pretty big it's got about 200 thousand organisations it's got about full 100 thousand notes and unlike most Web services all websites and it's got very spiky low profile 1 in the this this and his no website the so websites over the course of a day you've got like the little in the morning and then you Christmas Waco assessed grow slowly the hit the peak of the day's wage so be was like MIT definitely it's probably a P and then people go to bed it's us go back down again this is hosted chef request 2nd graph this is over 12 hours now you can see there but this spikes on the top of every hour and then smaller spikes and 30 minutes past the hour and cuts guess why that might be the anyone of chef go ahead John precisely the people of the main shaft just clients viet across he share fire hazard a future point as a demon then group it will actually pick a random time but if you put if I crumple the consult immediately on sometimes a little bit of splay and you can forget that we look for that and then a difference in the top what it means is that we have to cope with this huge increase in load that's appropriate and lots of under 15 minutes and 45 so yesterday I am your once the keynotes and you may have noticed and said hosted in 2010 and he mentioned this command 952 server creates and yet I'm talking here about moving chef to 80 US and not from a diverse social to explain all of that in 2010 hosted shaft actually he's to be addressed and then round
about 2011 we actually moved to and backspace by metal and because of performance issues the doesn't somewhat ironic that the performance issues are what led us to migrate back to enter the US last year so why do we actually move while we were Algren our our current structure we were of scaling issues boat these mean to address all these is that all you want me to the cloud cost and flexibility basically the ability to get what you want when you want it for as long as you want it and only pay for what you need an and ability scalable as needed so no lead time we after way treat them 2 weeks for the service for the it was 1 of the reasons develop time be so developers was spending a lot of the time supporting hosted and these and this is hosted will do nothing like and unconscious and so you use so all and it also and we'll be using the same configuration like chef cookbooks of recipes that we've been using the hosted since the time it in a tuple as previously so we still hard like configuration which is 0 . 9 and old ways of doing things and stuff modernize things so here it's a normal on crankshaft itself this is what many of you will install if you want to scale yo and checking structure 4 ships of you got low bounce the beginning they've got some from 10 1 or more and more if you need scale and then you've got a couple backends a hatred the this is not such a diagram for all cost the now you might not see very well that we've got the C 2 different low balances in genetics we got 3 islands is and anything different micro service that's an postage reference on the ships around and that that had its own machine or don't set machines so that's another the reason that we some and so the deaths would basically
spending way too much time supporting that big architecture viewed in deployments helping the troubleshooting and we want to build to make host chef look a lot more like the ships that's all you saw and if we can do that then we've get lots of benefits and we and we ops could ruin the hostage of initial chart we often to rely on that's provide acoustical parts because they provide the code for everybody else and this is how we get this is not such diagram from you hostage and it's got a few searches but out but overall it's pretty close to lot better than that
so we decided answer me we decided going to address and we decided to move the how willing to walk now we have tested in migration so it's but the migration of into about the mean different tasks all projects um beginnings learning at rest actually migrate the data and then monitoring if he was with them now my 2 was 2 weeks ago or so and the folks on the 1st the and so 1st task was set in the mid west the so you and packages and the package installed just co justices yeah we can figure that ships have starts the don't thank you it's not should that easy otherwise only here talking to you today and so 1 of the things we had to do was split house of the back services so I mentioned earlier that's true the and you had teachers to back and services in problem that phase you have and you stateful services so you've got ODS that's an that hold that's those all your shifted to the evils of search from at that age to scale you basically need to buy a bigger instance and so I don't see those when you explore on giant Godzilla instances we do what you want is a course like thousand an hour something ridiculous like that so withheld to our data and search in 2 separate sepsis so for the shifted best we said use and Bentley it's got the AMS as much as service you push a button database hand-delivered by magical Amazon elves the so that the self bits this this problem self Frank actually ODS would have really love our so far the chef database at least all this has been working very smoothly for the reporting database which I'll get to letter that's no sorry next month and backends
search and that has been the main components and they so surveys bad and Q which and so large hold all your such information we even make search right in q takes all the information when you make a change and and shuffling like the savior note it post onto the q and this set culture expander which takes the changes of the Q and then post a changes to cell and if you actually has never really brand new version Server and like you share a chair that's been released the past 2 weeks all of this is actually just Elastic Search and we what we decide you just keep only sepsis gather on the back a machine the and this will components that diagram but portal back and they're not matching gets and age database and that is each used by the front-end machines basically for things like feature flights web sessions and a way for us to enable maintenance sense what across all service you got genetics and the global answer the front and
the listening to various all checking by this fall information for things like help cement is not 4 that's we just used the less cash it sends basically hosted Metis as service and we shall use that but because of intense so on all posted and this is Paul to the reason for the diagram for you had all the services as shown by Frost of CID only its own server and mostly serves as a presurrealist so we wrote to combine these and even and normal shifts of installed some prompt all of these combined onto the same she's also the success just on from horizontal scale so this is a few all the poultry infection 1st off cookbook storage when you load your cookbooks the acts of physical data are and where you have your recipes and Putin's electrical better that it would twice 3 that make the migration process really easy because in old posted it was industry in new hosted was S 3 we have to do anything and all the moments Labonte infant everything which issues CELP and that was fairly decision I should have gone to the front-end slow but everything and then there's 1 of the machine we have hosted the e-mail me not have an in on current stock that is support box and basically that is a machine we can log into is completed the front-end but not the low and we use that fall troubleshooting password resets and any issues you make cross with Oak with many will fixes we use that spot and do all them on the spot box so the face sited now we have to get things went so 1st thing we do is put everything into also scaling groups now when he hit a autoscaling groups you think about machines match growing at you see the low-growing in new machines published house by the load that was thought he went to build a reality and write it down again and it takes about 10 minutes to spare you know so we don't actually use also scaling groups to also scouring from track we need be ample capacity anyway what we use a lot of skill is we set a static threshold and it will resonate show that there's always 16 note of 14 Note by we set it to and that means it is we both the idea that we want to treat all services couple and not that's greedy if we have a machines misbehave a weakness kilohertz and address with a spring can you change reply sets and I really let me get page so if we do not assume groups if you us all the scaling groups you need to build images and he's and to do that we use the we build image is every time we deploy change which isn't that often and let me show observations of my that all combi changes and when we do that we destroy all of our world and from tens our world um instances and so we we create new ones so most of the set up that we do it had to do in package build the power of the re the types of things that we do afterward so that is a this specific to the environment so anything that need to be different feature acceptance and delivered and all people proud and anything is specific to the instance itself things that host names of node names as specific a progression and credentials which there should have kind of a specific case of and and typical fake but we treated separately and because they can eat special that all additional consideration after brutes dementia the so that integration and when we had when she starts so we have used it a script that sets of shaft and segment onboard so they US even insects and a shell script that rooms on and on Beirut's when these comes up the and now we can't actually use the validated method that's the 1 where you don't need a valid key and good is we need she's too much themselves spots the office group so just drop validation key unit to share and tell shift to bone and she matched itself and it can and figure things and will see policy false much regression and that's been really useful for a number
of reasons and you can be about benefit oppose files and 1 big thing that we do a so we make use of so called Named lists and this is a separate room list for a note that you can ruin and by specifying by name so the what we do the impact we say use the build list and it just the bill technique and I'm going no woman list because all of the correction the shaft including the environment-specific credentials and the things mentioned earlier and so the actual edges US resources instances called yes and elastic cash we use turtle not meant is all of our address resources and it's made reaching really easy and you just have it's your infrastructure or other the instances and their full address resources as called and 1 way this is really helpful about old days before we migrated or the folders folder sketch of migrants we would happily settle in US West to 1 the public that most focused must look if the annuities 1 so we need we rigorous West here because of reasons boats and so on that decision to actually we was to US is want be close our customers are and we will to myself migration 18 ruefully 1 day and that migration that day with mostly transferring data that we plants and going through the tariff OK can check fake just removing any assumptions we have probably of our regions so I mentioned earlier when we do scaling groups we build an army and then we try our instances and recreate them problem is if you do that he try other instances Hausa just goes down so what we actually do days we rebuild our images and Parker we bring look new instances in sector autoscaling group that is not life and we were a few tests and then once we have been working with verified cell that we swap office in groups and while that lets us do it as a test beforehand and 8 and also lets us quickly go back things go wrong so we only do this for the front ends to and 1 reason for that is also still groups and will be points they were communion well stateless machines you can just killin instance and they'll come back and no 1 really cares both stable services such as back-end stuttering of search that has a lot data a meeting which will cover would that and we can't just kill instant because that might be the instance where the old data still so we just need to make sure that when we do and when we do our deployed meet change things we make sure we kill the secretary back and that's and then we fell power and them to be all primary and if you have single machines you can still use all the skill groups now he got single-gene chances out you don't really care about and about a going down because really it last in as instances randomly initialized the I the thing doesn't happen you got think it does the Book of Books I mentioned before some Xin Li sh into tuition troubleshooting alter reset passwords effect goes down for 5 minutes we don't boat he goes down 5 minutes and still 1 8 to be less to be being of replacement parts and that's happened numerous times global to just disappeared have a dynamic of this is a new box and I never even known because we don't need to monitor it we don't need to know because we don't really care the support boxes and they're onto which it separates so you can still use of celebrity single machines the so we note to the migration process itself we know how to and get implemented glass now he's migrants so like the part of the whole which but took the migration into a few different tasks means that that's the specifically the chefs that space and reported that stress leaving the search data and all the back-end data solar content and finally and actually flipping a switch that again it's the people over to the new and new the best thing at least a is migration the chefs so the chef such a database of Czech database server that is where all field that's still willing to know it's it's got an rules that it's got an you clients on the couple metadata basic everything except the coupled themselves and the In old hosted this was actually 3 databases we have 1 the shift of
itself then had to others phocids called by frost which gate of indication answers for all CID which also the authentication predicted service a new hosted we decided we should no jealous databases into 1 which is what I'm proudest of looks like so we chose on yes Elliot hosted at best 1 the that you can't do streaming application from post best into ODS the I think we might be able to now August an address it's got this like this this migration service and by the time we couldn't and we can do that we have to just copy the data and all the data wholesale from our old database into obvious and thankfully databases at no small all the data everyone in host a check actually so we will just copy everything across and D and cover the cost to a migration Nathan that list was proposed this this now reporting it is the ships over random checks over reported the and that he because of every shot from you you do events hosted which is a lot better the time migration about 2 to 3 terabytes it should also list so that's basically bigger boat that was something that we silicon just copy across in 5 minutes and all theUS on supposed to about 6 terabytes so some growth we could be an RTS so far reported as best we decided to use just knowledge his today's money post graphs which is meant we had to manage those databases boat it also meant we compute streaming replication and not Mingyuan migration File um revolted east we could just and replicates all data and then when the migration came we just fill that that space and migration is for that meeting which instance books before we could do replication unique database backup now old hosted that sensor at bare-metal image pogo we have to give a connection so the humor and comes out at this about 1 Terabyte every 3 hours by gigabit speeds now you never really get the full bandwidth for various reasons so where do to to be terabytes the let's allow full day the should put it out 3 years later the we made a few mistakes are collections this thing transferrin across the West so does field latency that slowdown transfer its let's see was the biggest issue we have because we went from Chicago so sprawling all across the US Middle country to the east coast and but we also assumed the we had full no more hosted traffic because well hosted we still had a certain and that it's about how local band relocation so where you don't how on top of that the potent that's stalls a and people querying as the Sheffer they've ever done at all a lease query Indonesia from 7 last month that's a lie data and reporting server disks were constantly saturated something that wasn't basically nothing left for was to transfer data across so that'll transfer rate we got like 10 megabytes sec 0 roughly 18 admits the and on top of that Our transfer to use Nelson to begin with that just died regularly like the transfer of copper is it would stop crash and and and then we have stuff going so we use a few tricks to try this speed things up we try policy think that some of these fancy tools you can get the claim to transfer your data 10 across the country 10 times faster than 6 capable of steel the latency issue and book all of these things the and it still took us of our week with those tricks to transfer them about 10 posters has these things called wall files which still replication data while you're taking the best backup we also so much In the time to this transfer from the wall files had protected and we could sort replication so I spent a we transfer and if they realize constant tension repeat process all over again so but to you letter we finally gets reporting transferred in the presence of reporting done
next service so so as image before solar holds all the search data and this is what it's queried when you when I search all in your recipes if you really such function against node of its data back to things like that base therefore data still works much smaller than the reported all best amount of a that solar has this built-in replication and it's not is not great application is not freedom application the better transfer the entire datasets spont every hour books it just forgot the dates that we have list would probably on the order of about 60 handy divides rather than 100 something the shoes they sold application and they just go across yeah and because they want to and wondering so shifts of CTL have also said just server rather it has a command called the index word it would take all the data that you have best and it will populate solar from the data and drafted its best we we didn't use that a is it just takes too long to do for this dataset size hosted will probably the least a day to read data from scratch so much easier to just transferred all the data as it is across from 1 to the other next but in the moment it is but it's only last catch that was even simpler the so he would call familiar that especially stored feature for X web sessions whether we in maintenance mode and all those things the data is pretty static and the data we had been old hosted that was speech a flock and settings for some ancient beach we added years ago so what we're going to do was just Stultz and the still the newly elastic cache of new voters that best with a completely fresh and blind at best we choose the settings that we needed and incrementally poets and you might recall sentence in for the host that green is deserve to migration itself and with size migrates at nite time it's this is really an cooled down time for was a check and units as full the all that 12 hour period and the Lotus fairly flat spikes boatman sensitive when at least most customers would not be looking but we started early index 1st task before it out lower your DNS details so frame doesn't know and TTL Indian is basically how long he do DNS lookup you'll laptop or your server activist in the local will keep the D. N. information we out looking up again which means if the person causing changes its such as to in migration use allele match so we loaded get to 60 seconds which meant when leaflets which most if not all of our customers and would change over within 60 seconds next we do test and but test migration and make few file combi changes and then later and we do our from migration process this thing but this the site the LP no wooden migrate the then we never makes as well and then we do we actually migrates that spaces that a of seizures mentioned fall flipping switches were we got replication on an initiation dish transferred will be don't of replication and and then using the so that picture the then we test you could test with an organization we've to countermand his mild measure and what then we you quit low test which we test million refers to a final low test may show an Indian Paul that we flip the unless both events is not posted and then we all that stuff site on migrated it so I mentioned blood test and that we made a homegrown tool culture swarm just simulates lots Mushaf and once and we used it to simulate peak-load on top of the all and then new hosted and we test against a single organization like just test OK we have ultimatums motor we still do it while everyone else still images mother the the problem with that best be very very careful about wording test against active Proc I also mentioned that's we tested single so we let me know this test no completely representative of a an an actual general so we expected that our low test widows if he's back and then we flip the switch and by interfere the high hat was before the switch real traffic come along and holding the ball it turns out the opposite happened you joint migration now we handle test earlier in the day we knew according to law test with good we or final test join migration just matrix is OK and everything both of hostages crash events and like what the heck is going on so we do which I call helps he shares with it OK we're not sure it's going on the Bowie thinking something moment test so we knew good to go but the switch and the was a foot switch like traffic and long and hosted took like traffic without any problem at all it turns out that when we did that will test join therefore maintenance we still had life traffic hitting the front-ends and then get animated smoke respond back at the same time you know all test what meant was that the engine next parts on the front ends they were basically taking all the group looking traffic was the tries and the test wants and right of so lesson learned that be very careful you load testing it do isn't always tell you the truth so we
migrated what is complete with DNS and it is great and these are the results and kind of see that both the top graph is right before the migration and that is the topolayout peak which at negatives it's entirely willfully 500 ms follow an Africa quest a method to the request right after the migration but went down to 50 ms clinical satisfactory 10 and just remind you that's all posted a big mess Walter surfaces hall to understand I'm not convinced I fully understood this architecture for migration problems of mostly will not convinced myself
and that's the oxygen afterwards much cleaner architecture which question which he's stunned architecture and we've got much improved upon a method that we got better build scale we can just add new instances we need and we're going the same chefs of factories as everyone here here is to ensure that tha the questions of something over here you because the so the question was always exposing hosted check host outside major grass and the answered yes we do have a lot of customers who locating restless 1 and some to some temporal migration we can check where most of us was made from the it's hosted shaft is available for anyone who wants the shaft and chances of if you were an group if you use the shift to begin with some light lunch at all and I shift tutorial you made it took shelter just when he really quickly and smes this question yeah 2 reports more harm in the theories of the is 0 or what your prior to we point front-ends acts a separate shifts so far and so regionally we have to and we just had such observer now and as of about approach mulch year will use in chapter 3 so that's a whole other talk this is the 1st work to each of the day or then quite possibly yes however when we do that I don't know what is the no so we we have to stick to the Masons window of about 4 hours and that was to allow time to transfer in all fire and police transferred over the whole shifted this and and also to allow time to compute Tunisia's can we probably have through the migration involves primal and the half the time so let's say 1 half hours and I'm a questions yes and that the total no count is about 400 thousand my notes and about 100 thousand yet you how stuff we have a team of about 4 or 5 people and so they the so we used to be on the temperature is actually what listen millipede big reason for migration in addition to the deaths and working on suggest we was spending all our time and deal with issues of shifting at the micro outages constantly to deal with these issues entities who migrated we have the bound to be not BookSampo stochastic so we have to accept it the FIL questions the OK well I give a much
Web Site
Subtraktion
Bit
Punkt
Gemeinsamer Speicher
Selbst organisierendes System
Gruppenkeim
Unrundheit
Ungerichteter Graph
Client
Web Services
Schnitt <Graphentheorie>
Gammafunktion
Verschiebungsoperator
Graph
Systemverwaltung
Profil <Aerodynamik>
Biprodukt
Last
Server
Skalarfeld
Dämon <Informatik>
Schlüsselverwaltung
Zeitzone
SaaS <Software>
Zentrische Streckung
Softwareentwickler
Subtraktion
n-Tupel
Natürliche Sprache
Code
Summengleichung
Arithmetisches Mittel
Virtuelle Maschine
Diagramm
Dienst <Informatik>
Skalierbarkeit
Menge
Mereologie
Computerarchitektur
Softwareentwickler
Datenstruktur
Konfigurationsraum
Streuungsdiagramm
Normalvektor
Bit
Mathematisierung
Versionsverwaltung
Dienst <Informatik>
Task
Virtuelle Maschine
CLI
Benutzerbeteiligung
Front-End <Software>
Migration <Informatik>
Zusammenhängender Graph
Elastische Deformation
Phasenumwandlung
Datenhaltung
Expandierender Graph
Natürliche Sprache
Softwarewartung
Dienst <Informatik>
Menge
Rechter Winkel
Ruhmasse
Server
Projektive Ebene
Bildschirmsymbol
Information
Verkehrsinformation
Instantiierung
Expandierender Graph
Quelle <Physik>
Webforum
Punkt
Prozess <Physik>
Momentenproblem
Nabel <Mathematik>
Adressraum
Gruppenkeim
Regulärer Graph
Benutzerfreundlichkeit
Extrempunkt
Raum-Zeit
Eins
Komponente <Software>
Metadaten
Client
Skalierbarkeit
Einheit <Mathematik>
Gruppentheorie
Lineare Regression
Skript <Programm>
Flächeninhalt
Tropfen
E-Mail
Figurierte Zahl
Verschiebungsoperator
Softwaretest
Zentrische Streckung
Addition
Schwellwertverfahren
Prozess <Informatik>
Datenhaltung
Gebäude <Mathematik>
Instantiierung
Dialekt
Hoax
Entscheidungstheorie
Teilmenge
Dienst <Informatik>
Datenfeld
Verbandstheorie
Menge
Server
Information
Normalspannung
Programmierumgebung
Schlüsselverwaltung
Instantiierung
Server
Stabilitätstheorie <Logik>
Subtraktion
Quader
Mathematisierung
Physikalismus
Zellularer Automat
Zahlenbereich
Gebäude <Mathematik>
Datenhaltung
Task
Virtuelle Maschine
Knotenmenge
Weg <Topologie>
Arithmetische Folge
Verkehrsinformation
Front-End <Software>
Migration <Informatik>
Datentyp
Luenberger-Beobachter
Passwort
Inhalt <Mathematik>
Speicher <Informatik>
Bildgebendes Verfahren
Hilfesystem
Demo <Programm>
Leistung <Physik>
Trennungsaxiom
Soundverarbeitung
Videospiel
Zehn
Diskretes System
Konfigurationsraum
Validität
Einfache Genauigkeit
Kanalkapazität
Schlussregel
Mailing-Liste
Elektronische Publikation
Integral
Office-Paket
Diagramm
Last
Softwareschwachstelle
Debugging
Turtle <Informatik>
Mereologie
Zentrische Streckung
Matrizenrechnung
Abstimmung <Frequenz>
Prozess <Physik>
Momentenproblem
Adressraum
Selbstrepräsentation
Gruppenkeim
Impuls
Wärmeübergang
Kartesische Koordinaten
Ungerichteter Graph
Extrempunkt
Gesetz <Physik>
Datensicherung
Lie-Gruppe
Raum-Zeit
Last
Softwaretest
Einheit <Mathematik>
Dämpfung
Gruppe <Mathematik>
Wärmeübergang
Datenreplikation
Randomisierung
Einflussgröße
Verschiebungsoperator
Softwaretest
Lineares Funktional
ATM
Krümmung
Datenhaltung
Stellenring
Abfrage
Bitrate
Frequenz
Ereignishorizont
Ultimatumspiel
Konstante
Softwarewartung
Dienst <Informatik>
Verknüpfungsglied
Datenfeld
Verbandstheorie
Menge
Automatische Indexierung
Rechter Winkel
ATM
Server
Information
Ordnung <Mathematik>
Lesen <Datenverarbeitung>
Instantiierung
Dualitätstheorie
Server
Explosion <Stochastik>
Web Site
Hecke-Operator
Selbst organisierendes System
Mathematisierung
Systemzusammenbruch
Sprachsynthese
Schar <Mathematik>
Datensicherung
Datenhaltung
Wurm <Informatik>
Task
Knotenmenge
Benutzerbeteiligung
Reelle Zahl
Verkehrsinformation
Front-End <Software>
Mini-Disc
Migration <Informatik>
Diskrete Simulation
Endogene Variable
Direkte numerische Simulation
Elastische Deformation
Indexberechnung
Bildgebendes Verfahren
Gammafunktion
Einfach zusammenhängender Raum
Videospiel
Matching <Graphentheorie>
Kegelschnitt
Zwei
Eindeutigkeit
Mailing-Liste
Elektronische Publikation
Caching
Mereologie
Debugging
Authentifikation
Wort <Informatik>
Bandmatrix
Lie-Gruppe
Resultante
Zentrische Streckung
Addition
Graph
Gruppenkeim
Zählen
Physikalische Theorie
Negative Zahl
Stochastik
Migration <Informatik>
Bildschirmfenster
Direkte numerische Simulation
Luenberger-Beobachter
Faktor <Algebra>
GRASS <Programm>
Computerarchitektur
Verkehrsinformation
Instantiierung
Verschiebungsoperator

Metadaten

Formale Metadaten

Titel It's Time for Us to Move: The Story and Migrating Hosted Chef to AWS
Serientitel ChefConf 2016
Autor Harrison, Mark
Lizenz CC-Namensnennung - Weitergabe unter gleichen Bedingungen 3.0 Unported:
Sie dürfen das Werk bzw. den Inhalt zu jedem legalen und nicht-kommerziellen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen und das Werk bzw. diesen Inhalt auch in veränderter Form nur unter den Bedingungen dieser Lizenz weitergeben.
DOI 10.5446/34639
Herausgeber Confreaks, LLC
Erscheinungsjahr 2016
Sprache Englisch

Inhaltliche Metadaten

Fachgebiet Informatik
Abstract Hosted Chef is one of the biggest Chef installations there is, with tens of thousands of organizations managing hundreds of thousands of Chef clients. By 2015, Hosted Chef had been growing exponentially for several years, and it was quickly outgrowing its home. It was time for a change, and so last October we migrated Hosted Chef from its original data center into AWS. As if the migration of a large production service wasn't enough, we were using an aging code base with practices and procedures that were years old, with references to CouchDB and workarounds from Chef 0.9! It was time to modernize all of our cookbooks, start using modern features, and generally rewrite everything at the same time. This talk is the story of that migration, the decisions we made, the challenges we faced, and the spectacular results. I'll cover what worked and what didn't go so well, and along the way I'll share some critical insights that will be useful to anyone running a large Chef installation in a cloud environment such as AWS.

Ähnliche Filme

Loading...
Feedback