Policy-based Cloud Storage: Persisting Data in a Multi-Site, Multi-Cloud World

© 2015 Apcera
Policy-based Cloud Storage
Persisting Data in a Multi-Site, Multi-Cloud World
V 2015-09-23.1
Earl C. Ruby III
Principal Software Engineer
Apcera
earl@apcera.com
@earlruby
http://earlruby.org

© 2015 Apcera
Introduction
My name is Earl, and I work at Apcera
This is the picture I used last year at RICON
Storage covers a lot of ground, so I’m going to
focus on file system storage and policy
Q&A afterwards, but feel free to ask questions

© 2015 Apcera
What problem are we trying to solve?

© 2015 Apcera
What problem are we trying to solve?
We want to reduce application
development time by allowing engineers
to easily provision services -- network,
DNS, NoSQL, DB, Web, etc. --- without
having to request the service from some
other group and without compromising
system security or stability

© 2015 Apcera
File systems in the cloud
! Provide a temporary file system to a single job
! Provide a persistent file system to a single job
! Provide a persistent file system shared with multiple jobs
Do all of the above on any cloud, anywhere

© 2015 Apcera
Temporary file system / single job
Not too interesting...
! Job starts, has a file system (how much data?)
! Job writes data to the file system (what kind of performance?)
! Data is stored locally on the same host where the job runs (limits the
available volume size)
! Job ends, data goes away (hence the name “temporary”)
! Linux containers handle this well today

© 2015 Apcera
Persistent file system / single job
More interesting...
! Job starts, has a file system
! Job writes data to the file system (what performance?)
! Job migrates to a new host, data moves with it (but how quickly?)
! Job restarts, data persists (how durable? RAID? 3 copy?)
! Job ends, data goes away (difference between restart and end?)

© 2015 Apcera
Persistent FS / shared / multiple jobs
Very interesting...
! First job starts, gets a file system (how much data?)
! More jobs start, can access same file system (how many jobs?)
! Jobs write data to the file system (how many jobs at the same time?
same file? same directory?)
! Jobs migrate to a new cloud (all jobs or some? does the shared data
migrate and if so, when?)
! Will it scale?

© 2015 Apcera
Define “Scale”
When we say “Scale”, what are we talking about?
! Total volume of data
! Number of simultaneous read/write/update/delete operations
! Number of simultaneous connections
! Total IOPS / Gbps (“Noisy Neighbor” problem)
! Consistency / Availability / Partition Tolerance (CAP)
! Predictable performance as all of the above increase
It depends...

© 2015 Apcera
Trade-offs
Different types of storage systems have different tradeoffs
! NFS - all traffic across network, tends to degrade if many jobs write to
the same file or directory (locking issues)
! Local SSD - fast, but total space is limited to the size of the disk, SPOF
! HDFS - Optimized for large files and sequential reads
! AWS EBS - Only works on AWS, can get expensive for high IOPS
! AWS Glacier - Cheap to store, slow and expensive to read. Optimized
for large files, write once, read rarely
! Legacy SAN - fast, works for on-premises cloud, not AWS, expensive
to maintain and extend

© 2015 Apcera
Policy!
Use policy for Provisioning,
Security, Performance, and
Business Logic.
Instead of trying to redesign the
wheel, describe what you want
to happen and let the application
platform make that happen.

© 2015 Apcera
Storage creation policy
! Max capacity (disk quota)
! Geo-replicated / HA fail-over required? (Y/N)
! Data Durability
○ Durability affects performance during recovery from hardware failure
○ Let the system decide if an app’s requirements means 1-copy, 2-
copy, 3-copy, RAID level, or erasure encoded
! Thick provisioning required? (Y/N)

© 2015 Apcera
Storage performance policy
! Min / Max IOPS (or best effort)
! Min / Max bandwidth
! Max Latency
! Max Concurrent Access
DO NOT define SSD / HDD / SAN -- define the performance you require,
let the platform figure out how to deliver that performance

© 2015 Apcera
Data handling policy
! Deduplication (Y/N)
! Compression (Y/N)
! At-rest encryption (None, LUKS, etc.)
! Point-in-time recovery required? (Y/N)
○ At what points ? (schedule)

© 2015 Apcera
Business policy
! Max Cost / GB
! Location - where is this data allowed to be physically located?
! Retention - keep data forever or delete after some date?

© 2015 Apcera
Play nice with others
In a multi-cloud world, your platform has to play nice with others
! Connect to other vendors’ storage solutions
! Be able to determine what policies those solutions support
! Apply policies that are supported by those solutions
! React gracefully when policies are not supported
! Self-heal if policy is not supported (or permitted) in any cloud -- give
the user a suggestion on how to move forward

© 2015 Apcera
Where are we now?
! We currently provide NFS volumes to Apcera jobs
! Provide persistent storage to Docker containers
! We are actively engaged with ClusterHQ/Flocker, ObjectiveFS, and
ConvergeIO, among others
! We are part of the Open Container Initiative
! Quotas and access permissions are supported by policy today

© 2015 Apcera
Policy example 1
The following is a basic
quota policy that limits
package size, RAM, disk,
and network resources for
the user Sam’s sandboxed
namespace. Without such
policy, Sam has an
unrestricted use of
resources.
quota::/sandbox/sam {
{ max.package.size 2GB }
{ total.package.size 6GB }
{ total.memory 5GB }
{ total.disk 15GB }
{ total.network 1Gbps }
}

© 2015 Apcera
Policy example 2
The following policy block
limits the maximum amount
of resources individual jobs
and job instances in the /
dev namespace may
consume.
quota::/dev {
{ max.job.cpu 200 }
{ max.instance.cpu 100 }
{ max.job.memory 64GB }
{ max.instance.memory 32GB }
{ max.job.disk 50TB }
{ max.instance.disk 25TB }
{ max.job.network 10Mbps }
{ max.instance.network 5Mbps }
}

© 2015 Apcera
Policy example 3
limits the total amount of
resources all jobs in the /
prod/website namespace
may consume.
quota::/prod/website {
{ total.cpu 1000 }
{ total.disk 10TB }
{ total.network 250Gbps }
}

© 2015 Apcera
Policy example 4
limits the maximum
memory and disk space all
job instances in the /prod
namespace may consume,
and the maximum memory
and disk space each job
instance in the /prod
namespace may consume.
quota::/prod {
{ max.instance.memory 256MB }
{ max.instance.disk 5GB }
{ total.disk 100GB }
}

© 2015 Apcera
When will we get there?
"WHERE ARE WE GOING?"
"PLANET TEN!"
"WHEN WILL WE GET THERE?"
"REAL SOON!"
-- Buckaroo Banzai

© 2015 Apcera
Thanks for listening!
Earl C. Ruby III
Principal Software Engineer
Apcera
earl@apcera.com
@earlruby
http://earlruby.org

Policy-based Cloud Storage: Persisting Data in a Multi-Site, Multi-Cloud World

Recommended

More Related Content

What's hot (20)

Similar to Policy-based Cloud Storage: Persisting Data in a Multi-Site, Multi-Cloud World (20)

More from Apcera (20)

Recently uploaded (20)

Policy-based Cloud Storage: Persisting Data in a Multi-Site, Multi-Cloud World