Browsed by
Tag: iaas

Going Cloud: the 8 don’ts

Going Cloud: the 8 don’ts

Okay, let’s face it: the world is finally figuring out that cloud is for everyone, and not just for large-scale enterprises. This is a big step ahead, but when it comes to new adopters there are still many misconceptions and wrong expectations.

Wrong expectations are probably the most common reason for failure, because they usually lead to disasters that leave moving back to a legacy infrastructure design as the only option left.

datacenter_scale

(Image Source: XKCD)

But, turns out, it’s easier than you would expect. There is a basic set of rules and guidelines, and if you follow them you can easily be successful.

Let’s begin in this article with the 8 don’ts:

  1. Never, ever trust a single instance of a given service. Don’t rely on redundant database platforms, replicated block devices, and so forth. They can still fail: accounting for this kind of failure at the application layer is the way to go.
  2. Don’t put all your eggs in a bucket: cloud platforms are available in different geographical locations by nature, so you should really leverage this. True geographical redundancy can be hard to achieve at the beginning, but try at least to have read replicas spread over the world, so that in case of downtime in the main region you’re using your service would just be degraded and not completely unreachable.
  3. Never think small. Some design patterns could seem overkills at first sight, but believe me, they are not. If you focus on designing your service so that it is ready for scaling up when needed, you won’t have to worry about later.
  4. Don’t design complex software platforms: micro services are the way to go. Keep them simple and easy to maintain. It will be easier to scale them, and not only from a technical point of view: imagine how easy could be handing over not a part of a complex software, but a micro service to a new dedicated development team.
  5. Never forget that performance is the key: a killer SQL query could still be affordable if you have a small number of users, but is going to be an issue when your platform grows. Make your application as efficient as possible, even when it doesn’t seem needed.
  6. Don’t forget that everything could break, at any time. Keep your instances as simple as possible, so that they are easy to operate. If one fails or starts misbehaving, just respawn it, don’t waste your time trying to fix or debug it. In an ideal world, they should all be stateless.
  7. Vertical scaling is a no go. Choose the size of your instances based on the performance you want a single request to have, but always spread multiple requests horizontally. This pattern will help a lot with availability as well.
  8. Don’t be ‘legacy’: the world around you is moving very fast, and just looking at it makes no sense. New releases of software packages usually improve their performance and efficienty, and new versions of the services your cloud provider is offering you usually improve a number of items, cost being usually the main one. Running a legacy instance type just because your platform is too hard to upgrade to a newer operating system makes no sense and will kill your business in the long term.

Here we are. Now go and build!

What an IaaS service is. And what it is not.

What an IaaS service is. And what it is not.

The term “Cloud Computing” has been openly used for almost ten years now, but there are still some misconceptions around the concept itself and around some more specific words like “IaaS” (Infrastructure as a Service).

Sometimes I have to face pointless discussions with people that have completely wrong ideas and expectations: this can be annoying from my point of view, but can be catastrophic for realities deciding to make “the big move” without having completely understood what the cloud is all about.

iaas

If you have come across this post as you’re still trying to figure out what “Cloud Computing” and “IaaS” mean, then let me save your life and probably your job with some clarifications.

The market offering isn’t helping us, as service providers are confused as well and they use to define “IaaS” completely unrelated products. The US NIST has released a document containing a list of 5 “Essential Characteristics” of cloud services, but they are not so specific and won’t help you make any choice.

When words are being used in such a confused way, you have to decide which of the many interpretations is the “authoritative” one: my authorities for this article are Amazon Web Services (and not because I work for them, but because ten years ago they have been the first at offering an IaaS platform) and OpenStack (that is, AWS concepts and terms reviewed by the biggest open source community in the cloud computing world).

So, what you should expect or not expect from an IaaS offering?

  • You should expect to be billed based on a Pay as you go model. Let’s be serious, if you have to pay an one time or monthly fee for your account and/or services you are using then this is not really cloud. Offering pay as you go services is a real technical challenge for the service provider, and if they aren’t giving you this option then you should have some doubts about them being up to date with the technology. Some providers will offer you discounts on long term commitments and this is fine, but always look for the PayG option, please.
  • You should expect to have full access to API and CLI tools and not just to a GUI. This is critical also if you are not planning to use them from the beginning. Cloud is all about automation, and if you stick with a service that only offers a GUI, then you will be forever bound to your mouse (and hands): if you come from an on premise physical server environment you could not see my point right now, but in the cloud you will start using automation soon, at least in its basic form. Because it’s easy and useful.
  • You should not expect your instances (virtual machines) to be always available. This is something I’ve already blogged about a few years ago (in italian, I’m sorry) but it’s still one of the biggest, most spread and more dangerous misconceptions. Cloud services are based on commodity hardware, and thus the instances on top of that should be considered in the same way, as a commodity. The single instance could be there or couldn’t be, and your customers don’t have to notice: you have to plan for high availability at application level, taking into account the various kinds of failure. Some additional services like Block Storage, Object Storage and Load Balancing as a Service will help you achieving the high levels of availability you need. If your service provider is offering you an extreme level of HA, then you’re probably paying for something you don’t need (if you’re using 5 web nodes, then what’s the matter if one of them goes down for a while?).
  • You should expect instant provisioning: seriously, provisioning has to happen in seconds. Be careful not to underestimate this: you could be happy with a 24 hours delivery time for your first bunch of servers, but believe me you won’t be when you will need to rapidly scale because of a traffic peak. Maybe I’m being too picky here but I expect the provisioning of my account to happen in real time as well: I’m not happy with providers asking me to send a physical signed contract or my IDs before using their service.
  • You should expect the service you choose not to have limits that could (and will) impact you. Okay, not all of us need the scale of AWS, but make sure your provider won’t go out of capacity when you will need it: planning for infrastructure is their job, and from your point of view you must always be able to use the resources you need, when you need them, with no previous commitment.
  • You should (probably) expect to have access to multiple autonomous regions: being it for active-active HA or just for backup and disaster recovery purposes, doesn’t make so much sense to choose a provider that is hosting its entire platform in a single datacenter. Yes, you could choose to use 2 different services providers hosting services in different locations, but this is not going to be easy to deal with.
  • You should (probably) expect not to be locked in by small-scale service providers: always look for open standards, expecially if the company you’re buying resources from is still at a scale where going out of business from one day to another is a (remote) possibility.
  • You should not expect to be able to easily scale vertically (increase instance size, or a single resource inside the instance): cloud computing is based on horizontal scalability (that means adding building blocks, not making the existing ones bigger), and this is why service provider don’t focus so much on hot resize of instances or on the ability to add RAM if you need RAM without modifying anything else. This is related to availability as well: if you can’t afford a planned downtime on a single instance in your infrastructure, then you’re doing something wrong.

That’s it, at least for now. I’m sure moving to the cloud is the right choice almost for every company in the world, but please make sure you fully understand it before making any choice. Really.

Giorgio

Design for failure

Design for failure

“Scusa ma non siamo capaci di offrirti la stabilità di cui hai bisogno, potresti pensarci tu?”

(Anonimo Cloud-Eretico sul Design for Failure)

C’erano una volta… gli sviluppatori, e le applicazioni. Gli sviluppatori si concentravano sul codice, dando per scontata la stabilità e la scalabilità dell’infrastruttura sottostante: usavano query SQL indescrivibili, ed era compito del sistemista farle girare velocemente, scrivevano codice senza gestione delle eccezioni perchè era compito del sistemista far si che quel determinato database server fosse sempre disponibile e non restituisse mai errori. Scrivevano software impossibile da distribuire su più macchine perchè tanto il sysadmin, in qualche modo, avrebbe fatto.

La colpa di ogni rallentamento o malfunzionamento di chi era? Del sistemista. Questo ha portato chi si occupa di infrastrutture a progettare soluzioni sempre più avanzate per far sopravvivere l’applicativo alle più inimmaginabili catastrofi, senza che questo subisse mai malfunzionamenti. Qualunque disgrazia fosse accaduta alle macchine che la servivano, l’applicazione sarebbe dovuta rimanere in piedi e funzionante.

Va detto che ci siamo (quasi) riusciti. Grazie alla virtualizzazione siamo arrivati a creare quello che è a tutti gli effetti hardware indistruttubile: la macchina fisica è diventata virtuale e quella virtuale sappiamo muoverla tra diversi nodi senza spegnerla, al solo costo di qualche millisecondo di freeze.

Abbiamo così creato piattaforme che astraevano quasi completamente la complessità sottostante, usando processori virtuali che restavano disponibili anche se quelli fisici prendevano fuoco e dischi virtuali che continuavano a servire dati anche se l’intero rack di storage veniva rubato dagli alieni.

Questa soluzione non era però ottimale: la replica sincrona, per esempio, era possibile solo in ristretti contesti geografici. Il costo di queste soluzioni era spesso proibitivo, e la loro complessità alta e non necessaria. Queste strutture, per quanto immortali potessero essere, erano sempre sotto la stessa autorità amministrativa. Tutto per non dare agli sviluppatori un compito in più: gestire la disponibilità dell’applicazione.

Screen Shot 2014-01-28 at 20.58.15

(Anonimo Cloud-Eretico che non ha compreso il ‘Design for Failure’)

Poi è arrivata una nuova generazione di developers: sviluppatori che volevano più controllo, volevano poter decidere come l’applicazione avrebbe reagito a malfunzionamenti dell’infrastruttura, e soprattutto si rifiutavano di pagare al fornitore complessi meccanismi di failover perchè… non ne avevano bisogno. Sapevano fare di meglio e sapevano farlo in modo più economico ma soprattutto più effettivo, più semplice.

Questi sviluppatori non chiedevano più a chi vendeva infrastrutture hardware immortale, chiedevano semplicemente del ferro: di qualunque tipo, prestazioni, forma, colore e dimensione, ed in ogni luogo. Si sarebbero occupati loro di inoltrare meno richieste ai processori meno potenti, di tenere in RAM i dati se i dischi della macchina erano troppo lenti. Si sarebbero curati loro di evitare di interrogare un database server che non rispondeva più ai comandi.

Volevano occuparsi, soprattutto, delle azioni di disaster recovery nel caso in cui un intero datacenter fosse andato a fuoco. Perchè nessuno meglio dello sviluppatore può sapere come deve reagire una applicazione a determinati eventi e di cosa questa ha bisogno.

Hanno poi iniziato a chiamarlo ‘Design for Failure’. La disponibilità non è più compito di chi gestisce l’infrastruttura: è l’applicazione ad esser progettata per far fronte a ogni evento o disgrazia, e la struttura sottostante fa solo il sollevamento pesi.

Nel modello ‘Design for Failure’ ognuno fa il suo lavoro: lo sviluppatore conosce l’applicazione e si occupa di farla funzionare, il gestore dell’infrastruttura si occupa delle prestazioni ma non si infila più in infiniti tunnel senza uscita per garantirne la disponibilità. Tutti risparmiano, perchè è tutto più semplice, con meno sovrapposizioni. Tutti vincono: perde solo chi non ha voglia di innovare.

Ecco perchè questo modello non è un fallimento, come tanti lo descrivono: è il futuro.