Azure

Why I Love Azure

• Deployment time
• [Usually static] Content stored in blob storage / CDN can be access directly by the client
• Maintaince and patching
• [Possibility] of dynamic scaling

Why I Hate Azure

• Deployment delays
• 1.3 diagnostics / start up sequence
• Throttling
• Billed for empty-except-for-continuation-token result sets, and for 503s

Top Azure Bugs

• Escape characters
• Buffer clearence
• Event order
• Other's HDD usage patterns can kill you
• Stuck in deployment mode
• [By design] Same queue message can be fetched by two threads at the same time
• Dequeue count doesn't apply to old queues

CDN

• DNS name resolves to local CDN
• 20 locations world wide, 99.95% uptime
• Cache expiry defaults to 72hours. Can be overridden in blob header (x-ms-blob-cache-control)
• SAS - Shared Access Signatures
• Put version in filename if want expiry-on-demand (e.g. logo-2011-02-27.png)

AppFabric Cache

• Same API as Windows Service AppFabric Cache
• Distributed caches scale out
• Subregion App Affinity
• AuthN/AuthZ with ACS (settings go in config)
• Integrates with ASP.NET Session state and Output Cache (TODO: How?)
• No object size limit
• Near cache (local client) for hot data (+ no serialization costs)
• Simple Get/Put syntax. Get returns NULL on cache miss

BLOBs

• Lease period 1 min. Lease easy way to determine master election
• 8k metadata
• ACLs directory & file level
• Each BLOB is in its own partition. If the BLOB is snapshotted, snapshots are in that partition too.
• Block
  • Two phase commit
  • Parallel out-of-order uploads
  • 200GB
  • Range reads possible
• Page
  • Range reads possible
  • Azure Drive
    • Up to 16 drives can be mounted
    • Drives are really page BLOBs mounted as a NTFS drive (hence 1TB limit)
    • Drives have write through caching

Queues

• Peek is 32 messages
• Use message count to determine need for more / fewer servers

Throughput

• Bandwidth: "a few hundred" MB/s
• Single Blob: 60 MB

Entity Groups

• Same table partition
• UP to 100 actions

ACS

• Security Token Service (STS)
  • Normalizes claims (can map (rename) claim names based on (amongst other things) providing organization. For example,: prefferedEMail => EMail
  • Intermediate Provider. Normal tokens from 3rd parties to 'Azure' format
• Move ASp.NET cache to resource disk for more space
• OR == table scan, use parallel queries

Misc / Perf

• IIS Warm-up extension
• Move ASp.NET cache to resource disk for more space
• OR == table scan, use parallel queries
• Max 5 upgrade domains
• Min 2 fault domains (e.g. different racks)
• Nodes not limited to 8 CPUs
• 3 VHDs
  • C:\ (Disk 1) = Temp DB
  • D:\ (Disk 0) = OS Image
  • E:\ (Disk 2) = Role
• Allocates
• Hard:
• # instances
• Hareware (CPUs / meme / etc)
• Fault Domains
• Update Domains
• Soft:
• Servicing of host OS / hypervisor (i.e. configurations that sit well with Fault Domains and Update Domains can be upgrade in parallel)
• Optimized network proximity between instances
• Host OS typically updated once a month
• FC host agent only allows access to IPs of VM on the same service, plus the internet at large.
• aaS's:
  • IaaS = Infrastructure as a Service. On-demand servers. Examples: Amazon EC2, VMWare vCloud
  • PaaS = Platform as a Service. On-demand hosting of services. Examples: Google AppEngine, Salesforce.com, Azure
  • Software as a Service. On-demand applications. Examples: G-Mail