Hadoop Security Design?Just Add Kerberos? Really?Andrew BechererBlack Hat USA 2010

Agenda Conclusion What is Hadoop Old School Hadoop Risks The New Approach to Security Concerns Alternative Strategies A Security Consultant Walks Into a Datacenter2

ConclusionDid Hadoop Get Safer?

ConclusionHadoop made significant advancesbut faces several significant challenges4

What is HadoopMapReduceSimplified ViewWho Is Using It

MapReduce Name Nodes & Data Nodes Data Access Job Tracker Job Submission Task Tracker Work Optional other services Workflow managers Bulk data distribution6

Simplified ViewUserJob TrackerTask TrackerTask TrackerTaskTaskHDFSHDFS7

Who is Using It8

Hadoop RisksInsufficient AuthenticationNo Privacy & No IntegrityArbitrary Code ExecutionExploit Scenario

Insufficient Authentication Hadoop did not authenticate users Hadoop did not authenticate services10

No Privacy & No Integrity Hadoop used insecure network transports Hadoop did not provide message level security11

Arbitrary Code Execution Malicious users could submit jobs which wouldexecute with the permissions of the Task Tracker12

Exploit Scenario Alice had access the Hadoop cluster Bob had access the Hadoop cluster Alice and Bob had to trust each other completely If Mallory got access to the cluster Alice and Bob bothdied in a fire.13

The New ApproachKerberosDelegation TokensNew Workflow ManagerStated Limitations

Kerberos Users authenticate to the edge of the cluster withKerberos (via GSSAPI) Users and group access is maintained in clusterspecific access control lists15

Delegation Tokens To prevent bottlenecks at the KDC Hadoop usesvarious tokens internally. Delegation Token Job Token Block Access Token SASL with a RPC Digest mechanism16

New Workflow Manager Oozie Users authenticate using some “pluggable”authentication mechanism Oozie is a superuser and able to communicate withJob Trackers and Name Nodes on behalf of the user.17

Stated Limitations Users cannot have administrator access to nodes inthe cluster HDFS will not transmit data over an untrustednetworks MapReduce will not transmit data over an untrustednetworks Security changes will not impact GridMixperformance by more than 3%.18

ConcernsQuality of Protection (QoP)Massive Scale Symmetric CryptographyPluggable Web UI AuthenticationIP Based Authentication

Quality of Protection (QoP)AuthenticationIntegrityPrivacy20

Symmetric Cryptography Block Access Tokens are used to access data TokenAuthenticator HMAC-SHA1(key, TokenID) The secret key must be shared between the NameNodes and all of the Data Nodes SHARED WITH ALL OF THE DATA NODES!!! That is alot of nodes.21

Pluggable Web UI Authentication There are multiple web Uis Oozie Job Tracker Task Tracker With no standard HTTP authentication mechanism Ihope your developers are up to it.22

IP Based Authentication HDFS proxies use the HSFTP protocol for bulk datatransfers HDFS proxies are authenticated by IP address23

Alternative StrategiesTahoe

Tahoe - A Least Authority File System Deserves its own talk Aaron Cordova gave one at Hadoop World NYC 2009 Disk is not trusted Network is not trusted Memory is trusted Intended for use in Infrastructure as a Service cloudcomputing environments Write performance is terrible but read performance isnot so bad25

Assessing HadoopTargetsTokens

Targets Oozie is a superuser capable of performing anyoperation as any user Name Nodes or Data Nodes can give access to all ofthe data stored in HDFS by obtaining the shared“secret key” Data may be transmitted over insecure transportsincluding HSFTP, FTP and HTTP Stealing the IP of an HDFS Proxy could allow one toextract large amounts of data quickly27

Tokens: Gotta Catch ‘em All Kerberos Ticket Granting Token Delegation Token Get the Shared Key if Possible Job Token Get the Shared Key if Possible Block Access Token Get the Shared Key if Possible28

Thank you for [email protected]