Note – the first post (Notes about the AWS JDK) can be found here
The following are the (minimally edited) notes I’ve taken while attending the session at AWS re:Invent conference in Las Vegas, NV on 27-30 November, 2012.
Keynote speech by Andy Jassy (Sr. Vice President, Amazon Web Services)
It is clear we are still at the ‘dawn age’ in cloud computing and PaaS in general: although AWS provides a rich API and lots of flexibility, if one draws a parallel with the evolution of programming languages, choosing which type of instance one should run one’s own webapp, is a bit like having to do ‘register allocation’ before compilers came round.
Considering the rise of mobile computing and the obvious fit between thin mobile client and cloud deployments, cloud-backed ‘Personal computing’ is the next trend.
There has been great emphasis (later even more strongly stated by Jeff Bezos) on the strategic difference between low-volume/high-margin and high-volume/low-margin model for the provision of public cloud services, with AWS firmly in the latter camp.
Reed Hastings (Netflix CEO) made an appearance and a testimonial about how AWS enable them to grow 1,000x over 4 years to 1BN hours/month watched streaming: this would have not even been thinkable on a self-owned data centre, let alone affordable.
Having access to a vast (and continually evolving) set of infrastructure services has enabled Netflix Engineering team to focus on ‘higher order’ problems and concentrate their efforts on their core business.
He also announced that ‘House of Cards’ w/ Kevin Spacey is coming out in Feb!
AWS is all about innovation
82 significant innovations introduced in 2012
158 new services
“Amazon Redshift is a fast and powerful, fully managed, petabyte-scale data warehouse service in the cloud. Amazon Redshift offers you fast query performance when analyzing virtually any size data set using the same SQL-based tools and business intelligence applications you use today.”
They have conducted an experiment with a subset of Amazon Retail data, 10x faster queries on 2BN rows
SAPSAP HANA One on AWS at $ 0.99/hr Dev Edition
Also ‘test drive’ to try it out
Integrated with other AWS services
Uses UI to drag-n-drop components
Connect logs generated to S3, EMR
It has scheduler and pre-conditions
They use the same ‘pipeline’ term
Can run ‘bash’ scripts stored in S3
Screenshot of the demo
Update: This has now been officially launched as the Data Pipeline, which is an exciting validation of SnapLogic‘s concept of integration flows as Snaps and Pipelines (see here)
… and low prices
Just announced an average 19% reduction on EC2 prices, and here at re:Invent Sassy just announced a 24-27% reduction on the price of S3 storage.
Key strategic objective is to facilitate Enterprise migration to cloud
Amazon VPC + Direct Connect + Route 53
Elastic Load Balancing
Dynamo DB is the fastest growing service in the history of AWS
Enterprise on AWS, no longer an ‘if’ but more a matter of ‘how far’
Java AWS SDK for Eclipse
Web UI frontend
EC2 back-end workers
S3 for the data storage
DynamoDB for metadata
SQS to distribute workload among workers
Suitable for CPU intensive apps, where processing must not be executed in the context of the Web UI
All this can be done via Eclipse AWS SDK
- Can manage multiple aws accounts
- Creates sample code working out of the box
- Explorer view to manage all supported services
In particular, you can see all S3 buckets, and explore its contents (including virtual directories); drag-and-drop works to/from the local filesystem
(note after upload, only owner can see — web ui can’t access: right click and set permission)
Explorer shows the tables on DynDb, create new tables, edit r/w capacity
DynDb editor shows just a page of results by default; can add a ‘scan condition’ to filter only a certain subset
Changes must be ‘saved’ to the actual db (using std ‘save’ cmd)
Can also execute remote debugging against a running instance.
(see my other post about Java SDK)
Can monitor and set thresholds for alarms/notifications
alerts can also take actions to scale up instances
Custom metrics (uses a ReST PUT API) — also can use scripts (search on G)
Stores 2-weeks worth of information, data can be pulled and pushed into long-term storage, analysis, map-reduce…
Asperatus (3rd party lib to push metrics)
Integrates with logging and JMX
Convention over configurarion
Reports metrics by instanceId/AppName
Can also use a Logger (reports class, err msg)
Easy JMX integration
Deployed on the front-end to measure ‘true’ latency perceived by clients
BigData with Spark/Shark
AMP Lab at Berkley Uni: htpp://amplab.cd.berkley.edu
see Mesos — cluster virtualization mgr (Twitter uses it for 2,500 VMs in prod)
BlinkDB — approximate querying system (ML)
Spark is a fast, distributed MapReduce-like engine (using in-memory storage)
General execution graphs
Supports Scala/Java (soon Python)
messages = spark.textfile(“hdfs://…”)
errors = messages.filter(_.startsWith(“ERROR”)
creates an RDD (resilient data distribution? )
Shark — port of Apache Hive to Spark
Compatible with existing Hive meta store and HDFS data
Dynamic join algo selection (done in real-time at query time)
About 100 times faster than Hive on Hadoop, even unstructured data
with GROUP BY (mandatory? ) takes a bit longer, but still 50x
CloudFront (AWS Content Distribution Network)
38 edge locations globally
When benchmarked for latency, CloudFront is 1st or 2nd v. Other CDN in US and other regions
Uses S3 for caching of content, with a Load Balancer in front, allows deployment to multiple Availability Zones (AZs) to offer HA.
Index service separate from Storage
Version Upgrade (the `new` way)
It used to be the case that, to upgrade a cluster of web servers, all sitting behind a LB, one would follow the usual pattern:
- light up a couple of ‘canaries’ with the new version and start routing some traffic to them; check for catastrophic (and not-so-catastrophic) failures, bugs, etc.
- once the ‘canaries’ survive, start deploying new instances with the new release and turn off the ones with the old – always keeping an eye on traffic, failures, latency, etc.
- if all goes well, one has managed a pretty serious upgrade cycle (with possibly thousands of nodes) without any downtime at all, and the users barely noticing a thing (if at all).
Awesom – the real problem is when, halfway through the upgrade, it turns out that a rollback is needed: at this point, you have a large number of servers, all in flux, without an easy way to tell which ones have the ‘new’ release, and which ones the ‘old’ – this could be a real operational nightmare.
Turns out that, at current prices, it’s a lot more cost-effective to just:
- again, use ‘canaries’ (if you are doing cloud deployments and don’t believe in staging, about time to start looking for a new job);
- leave the existing cluster alone, however, and just deploy an entire new one with the new release – once ready, just ‘flip the switch’ on your ELB and the new cluster starts serving traffic;
- anything serious happens, ‘flip the switch’ back, and you’re back in the stable configuration;
- rinse and repeat, until the new release is stable and serving traffic within the operational parameters (but, really, if you have to do this more than twice, it’s about time to find yourself a better dev team)
- Keep the ‘old version’ cluster around for a few days, just in case: the expenditure is minimal, and infinitely worth it if an emergency rollback is necessary.
Don’t make things more complex than they need to be
Design Architecture with your customers in mind, then use ‘late binding’ to pick which infrastructure serves you best
Deploying Python apps using BeanStalk
The python app can be developed as usual, without any change necessary.
source ./venv1/bin/activate (as usual)
To create the reqts use pip freeze >requirements.txt
pip -r requirements.txt
eb is a command-line tool to manage beanstalk
To create the env – generates the info to create it then use
to actually deploy and get it start.
to check status.
There is also an ini file that is read in ~/.elasticbeanstalk
EB uses also git integration to push the source/binary to the cloud:
Options can be found in: .elasticbeanstalk/optionsetting
It also allows to drive AWS AutoScaling triggers to start new EC2s (behind a LB); look for:
Overall it seemed like an interesting tool (the talk really went too fast to take any meaningful notes), I recommend reading the documentation here.