Touchless Security for Hadoop – combining API Security and Hadoop

It sounds like a parlor trick, but one of the benefits of API centric de-facto standards  such as REST and JSON is they allow relatively seamless communication between software systems.

This makes it possible to combine technologies to instantly bring out new capabilities. In particular I want to talk about how an API Security Gateway can improve the security posture of a Hadoop installation without having to actually modify Hadoop itself. Sounds too good to be true? Read on.

Hadoop Security and RESTful APIs

Hadoop is mostly a behind the firewall affair, and APIs are generally used for exposing data or capabilities for other systems, users or mobile devices. In the case of Hadoop there are three main RESTful APIs to talk about. This list isn’t exhaustive but it covers the main APIs.

  1. WebHDFS – Offers complete control over files and directories in HDFS
  2. HBase REST API – Offers access to insert, create, delete, single/multiple cell values
  3. HCatalog REST API – Provides job control for Map/Reduce, Pig and Hive as well as to access and manipulate HCatalog DDL data

These APIs are very useful because anyone with an HTTP client can potentially manipulate data in Hadoop. This, of course, is like using a knife all-blade – it’s very easy to cut yourself. To take an example, WebHDFS allows RESTful calls for directory listings, creating new directories and files, as well as file deletion. Worse,  the default security model requires nothing more than inserting “root” into the HTTP call.

To its credit, most distributions of Hadoop also offer Kerberos SPNEGO authentication, but additional work is needed to support other types of authentication and authorization schemes, and not all REST calls that expose sensitive data (such as a list of files) are secured. Here are some of the other challenges:

  • Fragmented Enforcement – Some REST calls leak information and require no credentials
  • Developer Centric Interfaces – Full Java stack traces are passed back to callers, leaking system details
  • Resource Protection – The Namenode is a single point of failure and excessive WebHDFS activity may threaten the cluster
  • Consistent Security Policy – All APIs in Hadoop must be independently configured, managed and audited over time

This list is just a start, and to be fair, Hadoop is still evolving. We expect things to get better over time, but for Enterprises to unlock value from their “Big Data” projects now, they can’t afford to wait until security is perfect.

One model used in other domains is an API Gateway or proxy that sits between the Hadoop cluster and the client. Using this model, the cluster only trusts calls from the gateway and all potential API callers are forced to use the gateway. Further, the gateway capabilities are rich enough and expressive enough to perform the full depth and breadth of security for REST calls from authentication to message level security, tokenization, throttling, denial of service protection, attack protection and data translation. Even better, this provides a safe and effective way to expose Hadoop to mobile devices without worrying about performance, scalability and security.  Here is the conceptual picture:

 

Intel(R) Expressway API Manager and Intel Distribution of Apache Hadoop

In the previous diagram we are showing the Intel(R) Expressway API Manager acting as a proxy for WebHDFS, HBase and HCatalog APIs exposed from Intel’s Hadoop distribution. API Manager exposes RESTful APIs and also provides an out of the box subscription to Mashery to help evangelize APIs among a community of developers.

All of the policy enforcement is done at the HTTP layer by the gateway and the security administrator is free to rewrite the API to be more user friendly to the caller and the gateway will take care of mapping and rewriting the REST call to the format supported by Hadoop. In short, this model lets you provide instant Enterprise security for a good chunk of Hadoop capabilities without having to add a plug-in, additional code or a special distribution of Hadoop. So… just what can you do without touching Hadoop? To take WebHDFS as an example the following is possible with some configuration on the gateway itself:

  1. A gateway can lock-down the standard WebHDFS REST API and allow access only for specific users based on an Enterprise identity that may be stored in LDAP, Active Directory, Oracle, Siteminder, IBM or Relational Databases.
  2. A gateway provides additional authentication methods such as X.509 certificates with CRL and OCSP checking, OAuth token handling, API keys support, WS-Security and SSL termination & acceleration for WebHDFS API calls. The gateway can expose secure versions of the WebDHFS API for external access
  3. A gateway can improve on the security model used by WebHDFS which carries identities in HTTP query parameters, which are more susceptible to credential leakage compared to a security model based on HTTP headers. The gateway can expose a variant of the WebHDFS API that expects credentials in the HTTP header and seamlessly maps this to the WebHDFS internal format
  4. The gateway workflow engine can maps a single function REST call into multiple WebHDFS calls. For example, the WebHDFS REST API requires two separate HTTP calls for file creation and file upload. The gateway can expose a single API for this that handles the sequential execution and error handling, exposing a single function to the user
  5. The gateway can strip and redact Java exception traces carried in the WebHDFS REST API responses ( for instance, JSON responses may carry org.apache.hadoop.security.AccessControlException.* which can spill details beneficial to an attacker
  6. The gateway can throttle and rate shape WebHDFS REST requests which can protect the Hadoop cluster from resource consumption from excessive HDFS writes, open file handles and excessive  create, read, update and delete operations which might impact a running job.

This list is just the start, API manager can also perform selective encryption and data protection (such as PCI tokenization or PII format preserving encryption) on data as it is inserted or deleted from the Hadoop cluster, all by sitting in-between the caller and the cluster. So the parlor trick here is really moving the problem from trying to secure hadoop from the inside out to moving and centralizing security to the enforcement point. If you are looking for a way to expose “Big Data” outside the cluster, an the API Gateway model may be worth some investigation!

Blake

 

 

Blake Dournaee

About Blake Dournaee

Passionate, energetic product manager that lives at the intersection of business, innovation and technology. I currently work at Intel in the Datacenter Software Division working on products and technologies relating to mobile, APIs, cloud services and big data.

One Response to Touchless Security for Hadoop – combining API Security and Hadoop