Ozone- Object store for Apache Hadoop

1 © Hortonworks Inc. 2011 – 2016. All Rights Reserved
Ozone – Object Store for Apache
Hadoop
Anu Engineer
aengineer@apache.org
Arpit Agarwal
arp@apache.org

Ozone – Why an Object Store
⬢ With workloads like IoT we are looking at scaling to trillions of objects.
⬢ Apache HDFS is designed for large objects – not for many small objects
– Small files create memory pressure on namenode.
⬢ Each small file creates a block in the datanode.
–Datanodes send all block information to namenode in BlockReports.
⬢ Both of these create scalability issues on Namenode.
⬢ Metadata in memory is the strength of the original GFS and HDFS design, but also its
weakness in scaling number of files and blocks.
⬢ An object store has simpler semantics than a file system and is easier to scale
Apache Hadoop, Hadoop, Apache are either registered trademarks or trademarks of the Apache Software Foundation in the United States and other countries.

Ozone – Why an Object Store (continued)
⬢ Ozone attempts to scale to trillions of objects
– This presentation is about how we will get there.
⬢ Ozone is built on a distributed metadata store.
⬢ Avoids any single server becoming a bottleneck
⬢ More parallelism possible in both data and metadata operations
⬢ Build on well tested components and understood protocols
–RAFT for consensus
•RAFT is a protocol for reaching consensus between a set of machines in an unreliable
environment where machines and network may fail.
–Off-the-shelf Key-Value store like LevelDB
•LevelDB is an open-source standalone key-value store built by Google.

Alternative solutions to NameNode scalability
⬢ HDFS federation aims to address namespace and Block Space scalability issues.
–Federation deployment and planning adds complexity
–Requires changes to other components in the Hadoop stack
⬢ HDFS-8286 - Partial Namespace In Memory.
–Proposal to keep active working set of namespace in memory.
⬢ HDFS-5477 - Block Management as a Service.
–Proposed solution for block space scalability issue.
⬢ Ozone borrows many ideas from these and is a super set of these approaches.

5 © Hortonworks Inc. 2011 – 2016. All Rights Reserved5 © Hortonworks Inc. 2011 – 2016. All Rights Reserved
Presentation Outline
Ozone Introduction
Ozone Architectural Overview
Containers
Ozone - Bringing it all together
Bonus Slides - if we have time.

Ozone - Introduction
⬢ An Ozone URL
–http://hostname/myvolume/mybucket/mykey
⬢ An S3 URL
–http://hostname/mybucket/mykey
⬢ An Azure URL
–http://hostname/myaccount/mybucket/key

Ozone - Definitions
⬢ Storage Volume
–A notion similar to an account
–Allows admin controls on usage of the object store e.g. storage quota
–Created and managed by admins only
⬢ Bucket
–Consists of keys and objects
–Similar to a bucket in S3 or a container in Azure
–ACLs

Ozone - Definitions (continued)
⬢ Storage Key
–Unique in a bucket
⬢ Object
–Values in a bucket
–Each corresponds to a unique key within a bucket

Ozone - REST API
⬢ POST - Creates Volumes and Buckets
–Only Admin creates volumes
–Bucket can be created by owner of the volume
⬢ PUT - Updates Volumes , Buckets and creates keys
–Only admin can change some volume settings
–Buckets have ACLs
–Creates Keys

1
0
© Hortonworks Inc. 2011 – 2016. All Rights Reserved
Ozone - REST API (continued)
⬢ GET - Lists volumes and buckets and allows reading of keys
–Lists Volumes
–List Buckets
–Get Keys
⬢ DELETE - Deletes volumes, buckets and keys.
–Delete Volumes
–Delete Buckets
–Removes the Key

1
1
Ozone Components
⬢ Containers – Actual storage locations on Datanodes.
–We acknowledge the term container is overloaded. No relation to YARN containers or LXC.
–Assume container means a collection of blocks on a datanode for now.
–Containers deep dive to follow.
⬢ Ozone Handler - REST frontend for the Ozone rest protocol - deployed on datanodes.
⬢ Storage Container Manager (SCM) - Manages the container life cycle.
⬢ Ozone Key Space Manager (KSM) - Maps Ozone entities to Containers.
⬢ Container Client - Talks to KSM to discover the location of a container and sends IO
requests to the appropriate container.

1
2
Ozone Overview

1
3
Ozone Key Space Manager

1
4
⬢ Key to container mapping service.
⬢ Keeps the key ranges to containers mapping in memory.
–Θ(number of containers) - 1 Exabyte cluster = 200M containers x 5GB each.
–Memory usage scales with number of containers and not number of keys.
⬢ KSM does NOT know about all the keys in the system.
⬢ KSM state is replicated via RAFT, NameNode-like snapshots.

1
5
⬢ KSM knows about Ozone Volumes and Buckets.
⬢ KSM keeps a map of Volumes to container and buckets to containers.
⬢ KSM performs longest prefix match on a given string.
⬢ Example: The user wants to lookup a key - “/volume1/bucket1/key1”
–KSM authenticates the user, maps this key to a container and looks up the container location.
–Container client gets a token from the KSM and talks to the container on the data node.
–Container client makes a getKey call to the datanode container with the full key path.
–DataNode validates the access token and serves the value.
⬢ Contents of a bucket may span multiple containers.

1
6
KSM - Bucket spanning multiple containers

1
7
© Hortonworks Inc. 2011 – 2016. All Rights Reserved1
7
Containers

1
8
Container Framework
⬢ A shareable generic block service that can be used by distributed storage services.
⬢ Make it easier to develop new storage services - BYO storage format and naming
scheme.
⬢ Design Goals
–No single points of failure. All services are replicated.
–Avoid bottlenecks
•Minimize central state
•No verbose Block Reports
⬢ Lessons learned from large scale HDFS clusters.
⬢ Ideas from earlier proposals in HDFS community.

1
9
Container Framework Components

2
0
Containers

2
1
Containers
⬢ A container is the unit of replication
–Size bounded by how quickly it can be re-replicated after a loss.
⬢ Each container is an independent key-value store.
–No requirements on the structure or format of keys/values.
–Keys are unique only within a container.
⬢ E.g. key-value pair could be one of
–An Ozone Key-Value pair
–An HDFS block ID and block contents
•Or part of a block, when a block spans containers.

2
2
Containers (continued)
⬢ Each container has metadata
– Metadata consistency maintained via the RAFT protocol
– Metadata consists of keys and references to chunks.
⬢ Container metadata stored in LevelDB.
– Exact choice of KV store unimportant. LevelDB is already used by other Hadoop components.
⬢ A chunk is a piece of user data.
– Chunks are replicated via a data pipeline.
– Chunks can be of arbitrary sizes e.g. a few KB to GBs.
– Each chunk reference is a (file, offset, length) triplet.
⬢ Containers may garbage collect unreferenced chunks.
⬢ Each container independently decides how to map chunks to files
– Allow reauthoring files for performance, compaction and overwrites.

2
3
Containers (continued)

2
4
Containers support simple client operations
⬢ Write chunks - streaming writes
⬢ Put(key, List<ChunkReference>)
–The value is a list of chunk references.
–Putting a key makes previously written chunk data visible to readers.
–Put overwrites previous versions of the key.
⬢ Get(key)
–Returns a list of chunk references
⬢ Read chunks - streaming reads
⬢ Delete(key)
⬢ List Keys

2
5
Storage Container Manager

2
6
⬢ A fault-tolerant replicated service
⬢ Replicates its own state using RAFT protocol
⬢ Provides Container Location Service to clients
–Given a container ID, return a list of nodes with replicas
–Mapping a container ID to Data Nodes (and vice versa)
⬢ Provides Cluster Membership Management
–Maintain list of live Data Nodes in the cluster
–Handle heartbeats from DataNodes

2
7
Storage Container Manager (continued)
⬢ Provides Replication services
–Detect lost container replicas and initiate re-replication
–Containers send a container report.
•Unlike HDFS block reports which include details about each block , a container report is a
summary of information.
•This is used by KSM for placement of containers
⬢ If a node suffers from disk failure or if a node is lost, the reconstruction is a local
activity which is coordinated via RAFT running on the data nodes.

2
8
⬢ Maintains pre-created containers
⬢ Collects container operation statistics
⬢ Decides which Data Nodes form the replication set for a given container.
–The number of replication sets in a cluster is bounded
–Borrowing the work done by Facebook and RAMCloud (Copysets, Cidon et al. 2013).
⬢ Important - Knows nothing about keys
–Does NOT provide Naming Service (mapping keys to containers)
–e.g. KSM provides naming for Ozone.

2
9
Conceptual Representation of Ozone and Container State

3
0
0
Ozone - Bringing it all together

3
1
Bringing it all together - Ozone createVolume operations
Key Space Manager
Replicated containers
1: createVolume
Container Manager
2: Lookup(volName, Operation) 3: getContainer
4: putMetdata(VolumeName, Properties)
Ozone HandlerClient
Heartbeats & Reports
DataNodes

3
2
Tracing an Ozone PutKey
Key Space Manager
Replicated containers
1: Ozone - putKey
2: Lookup(keyName, Operation)
4: putData(File, offset, Length, data)
Client
5: putMetadata(key, List<chunks>)
Ozone Handler
DataNodes

3
3
Tracing an Ozone createVolume
OzoneVolume vol =
(new VolumeBuilder(pipeLine))
.setCreated(new Date())
.setOwnerName("bilbo")
.setClient(client)
.setName(“shire”)
.build();
POST /shire
keyData = {ContainerKeyData}
keyName = "shire"
containerName = "OzoneContainer"
metadata =
0."Created" -> "1449533074362"
1."CreatedBy" -> "gandalf"
2."Key" -> "VOLUME"
3."Owner" -> "bilbo"
Ozone REST Handler code Container wire and storage format

3
4
Ozone Metadata operations
⬢ Any metadata write to a container is Replicated via RAFT.
⬢ Machines forming the replication set for a container comprise a pipeline.
⬢ A createVolume call reduces to putKey operation on the container.
⬢ putKey is consistent, atomic and durable.
⬢ All metadata data operations are done via putKey, getKey and deleteKey.
⬢ Data is written to one or more chunks and a key is updated to point to those chunks.
⬢ Updating the key makes the data visible in the namespace.

3
5
Current State of Ozone
⬢ Stand alone container framework.
⬢ Single node ozone using container framework.
⬢ Full REST API -- Command Line Tools and Client Libs are fully functional.
⬢ Active development in branch HDFS-7240.
⬢ Work in progress:
–SCM
–KSM
–Replication Pipeline
–RAFT

3
6
Acknowledgements
⬢ Ozone is being designed and developed by Jitendra Pandey, Chris Nauroth, Tsz Wo
(Nicholas) Sze, Jing Zhao, Suresh Srinivas, Sanjay Radia, Anu Engineer and Arpit
Agarwal.
⬢ The Apache community has been very helpful and we were supported by comments
and contributions from Kanaka Kumar Avvaru, Edward Bortnikov, Thomas Demoor,
Nick Dimiduk, Chris Douglas, Jian Fang, Lars Francke, Gautam Hegde, Lars Hofhansl,
Jakob Homan, Virajith Jalaparti, Charles Lamb, Steve Loughran, Haohui Mai, Colin
Patrick McCabe, Aaron Myers, Owen O’Malley, Liam Slusser, Jeff Sogolov, Enis
Soztutar, Andrew Wang, Fengdong Yu, Zhe Zhang & khanderao.

3
7
7
Thank You

3
8
8
Bonus Slides

3
9
Ozone Key Space Manager - Dynamic Container Partitioning
⬢ KSM deals with dynamic partitioning of containers.
⬢ Let us say that a user starts by uploading all his photographs to a bucket in ozone.
⬢ Since all the photographs are called IMG_* (thanks Apple), we will soon overflow the
5GB capacity of the container.
⬢ At this point we need to split the container.

4
0
Ozone KSM - Dynamic Container Partitioning
⬢ The container client attempts to write the Nth ozone key, IMG_N, gets a partition
required error.
⬢ Container client will take that error and return that info to KSM.
⬢ That error contains the info -- about the proposed split -- That is IMG_0- IMG_200 will
stay in this container and IMG_201-IMG_400 will move to next container.
⬢ Note: KSM initiates container partitioning but mechanics of the split are handled by
the Container Layer

4
1
⬢ One of the assumptions we have made about a container split is that the splits are on
the same datanode as the original container.
⬢ This allows us to reduce a split operation to a copy of Keys from one LevelDB to
another LevelDB.
⬢ if we need to move actual file data from one datanode to another -- we do support
container moves. However they are slow.
⬢ A split on the other hand will complete in seconds in most cases.
⬢ The split point is chosen by the container so that we are able to pick the 50th percentile
position that gives us reasonable chance at an equal partition of a container.
⬢ KSM does not know about the keys or the actual data sizes until much later.
⬢ So always relies on the container to tell it where the split should be.

4
2
⬢ A container split is done in KSM via updating the Tree. The range partition key we
maintain gets updated to reflect the fact that Keys - {IMG_0 - IMG_200} are on
container C1, and keys {IMG_201-IMG_Z} are on C2.
⬢ Container will update the SCM when the split is done.
⬢ This information is learned and maintained by KSM.

4
3
Ozone Key Space Manager - Soft Quotas
⬢ In the HDFS world, a Quota is hard limit. It is actually conservative in quota
management.
⬢ In the ozone world, Quotas are soft quotas. That is users can and will be able to violate
it, but KSM/SCM will eventually learn about it and lock the volume out.
⬢ The key reason why this is different is because KSM/SCM is not involved in the
allocation of chunks.
⬢ The containers have a partial -- that is an isolated view of the data in a volume. Since
volumes can span many containers, it is possible for users to allocate chunks that
violate the volume quota, but eventually KSM will learn and disable further writes to a
volume.

4
4
Ozone Key Space Manager - Missing namespace problem
⬢ One great thing about HDFS is Namenode.
–Despite scalability issues, in most cases Namenode does a wonderful job.
⬢ Ozone - Subtle problem if we lose the all replicas of a container.
⬢ We will not only lose data -- just as if HDFS lost its all 3 replicas, but we will also lose
information about which keys have been lost.
⬢ To solve this issue, we propose to have a on-disk eventually consistent log maintained
by a separate service.
–Records information about the keys that exist in the cluster.
⬢ This Scribe service logs the state of the cluster.

4
5
Ozone - Range reads
⬢ Ozone supports range reads and might support range writes like part upload in S3.
–Ozone achieves this by using the chunk mechanism.
–Chunks offer a stream like interface.
–You can seek to any location and read as many bytes as you want.
–This is used by ozone to support range reads
⬢ Periodic Scanner can reclaim unreferenced chunks.

4
6
Scalability - What HDFS does well
⬢ HDFS NN stores all namespace metadata in memory (as per GFS)
–Scales to large clusters (5K) since all metadata in memory
•60K-100K tasks can share the Namenode
•Low latency
–Large data if files are large
⬢ Proof points of large data and large clusters
–Single Organizations have over 600PB in HDFS
–Single clusters with over 200PB using federation
–Large clusters of over 4K multi-core nodes hitting a single NN

Ozone- Object store for Apache Hadoop

In this document

More Related Content

What's hot

Similar to Ozone- Object store for Apache Hadoop

More from Hortonworks

Recently uploaded

Ozone- Object store for Apache Hadoop