How do you envision things looking in the intermediate phase where we have support for both security modes?
I imagine it's easy enough on the Shiro side of if we go with the AOP annotations for authorization (the interceptor can just check if security_mode == SHIRO before doing anything), but that means we'd still have the legacy sessionValidator code in every RPC impl that would need to be wrapped in the inverse check (security_mode == CAPABILITY_VALIDATOR). Would it make sense to do a first pass to refactor the existing auth checking logic to a reusable method, or are we ok living with the temporary ugliness involved in adding that mode checking wrapper to all the existing auth code? On Wed, Feb 11, 2015 at 3:23 PM, Kevin Sweeney <kevi...@apache.org> wrote: > On Wed, Feb 11, 2015 at 3:19 PM, Zameer Manji <zma...@apache.org> wrote: > > > +1 to this proposal. > > > > Will we have dual implementations of API methods as we deprecate the > > SessionKey based API methods? > > > Yes for backwards-compatibility I think we'll need a flag to indicate which > system to use. It will probably be an all-or-nothing setting (think > -security_mode=SHIRO|CAPABILITY_VALIDATOR). > > > > > On Wed, Feb 11, 2015 at 3:13 PM, Kevin Sweeney <kevi...@apache.org> > wrote: > > > > > I've been thinking about revamping the authentication and authorization > > in > > > the scheduler recently. I've investigated Apache Shiro > > > <http://shiro.apache.org/> and I think it would fit into the scheduler > > > nicely as a replacement for our custom CapabilityValidator > > > < > > > > > > http://people.apache.org/~kevints/aurora/dist/0.5.0-incubating/javadoc/org/apache/aurora/auth/CapabilityValidator.html > > > > > > > framework (for which there currently exists no implementation). > > > > > > I'd like feedback on this proposal. > > > Status Quo > > > > > > Security is currently implemented by a hand-rolled SessionValidator > > > < > > > > > > http://people.apache.org/~kevints/aurora/dist/0.5.0-incubating/javadoc/org/apache/aurora/auth/SessionValidator.html > > > > > > > framework. No public implementations exist. > > > Proposal > > > > > > Change the scheduler to use the Apache Shiro framework for > authentication > > > and authorization. Move authentication from application to transport > > layer > > > and move authorization to the Shiro Permissions model. > > > Advantages > > > > > > A few things that will become possible once this work is complete: > > > > > > 1. Ability to configure secure Aurora client-to-scheduler with a simple > > > flat configuration file (shiro.ini > > > <http://shiro.apache.org/configuration.html>). > > > > > > 2. Ability to integrate Aurora with my enterprise SSO (Kerberos+LDAP > for > > > example) by implementing a custom Shiro Realm > > > <http://shiro.apache.org/realm.html>. > > > > > > 3. Ability to allow a CI server to continuously deploy to every role's > > > "staging" environment without being able to touch its "prod" one by > using > > > Shiro's WildcardPermission > > > < > > > > > > https://shiro.apache.org/static/1.2.3/apidocs/org/apache/shiro/authz/permission/WildcardPermission.html > > > > > > > . > > > > > > 4. Ability to authenticate to the scheduler API using Kerberos (via > > SPNEGO > > > <http://spnego.sourceforge.net/>) or HTTP Basic auth. > > > > > > 5. Ability to perform authenticated write operations on a job via the > web > > > UI > > > < > http://www.chromium.org/developers/design-documents/http-authentication > > >. > > > Suggested Reading > > > > > > Shiro has excellent documentation and is a fellow Apache Foundation > > > project. I suggest you check out at least the 10-minute tutorial > > > <http://shiro.apache.org/10-minute-tutorial.html> and the Guice > > > integration > > > documentation <http://shiro.apache.org/guice.html>. > > > Scheduler-side changes > > > > > > The best way to show the proposed changes is by example. In addition to > > > Guice wiring changes to place the Shiro authentication filter into the > > > request chain, code that previously looked like > > > > > > @Override > > > > > > public Response createJob( > > > > > > JobConfiguration mutableJob, > > > > > > @Nullable final Lock mutableLock, > > > > > > SessionKey session) { > > > > > > requireNonNull(session); > > > > > > try { > > > > > > sessionValidator.checkAuthenticated( > > > > > > session, > > > > > > ImmutableSet.of(mutableJob.getKey().getRole())); > > > > > > } catch (AuthFailedException e) { > > > > > > return errorResponse(AUTH_FAILED, e); > > > > > > } > > > > > > // Request is authenticated and authorized, continue. > > > > > > } > > > > > > becomes > > > > > > @Override > > > > > > public Response createJob( > > > > > > JobConfiguration mutableJob, > > > > > > @Nullable final Lock mutableLock) { > > > > > > // subject is injected in the constructor by Guice each request. > > > > > > // checkPermission will throw an unchecked > > > > > > // AuthorizationException that bubbles up as a 401. > > > > > > // This line could also be inserted by inspection of the method > > > > > > // call in a security AOP layer. > > > > > > subject.checkPermission( > > > > > > // A Shiro WildcardPermission job:create:mesos:prod:labrat > > > > > > new JobScopedPermission("job:create", mutableJob.getKey())); > > > > > > // Request is authenticated and authorized, continue. > > > > > > } > > > > > > Some admin methods are protected by annotations like > > > > > > @Requires(Capability.PROVISIONER) > > > > > > public Response startMaintenance(Set<String> hosts, SessionKey session) > > { … > > > } > > > > > > They'd become > > > > > > @RequiresPermission("maintenance:create") > > > > > > public Response startMaintenance(Set<String> hosts) { … } > > > Client-side changes > > > > > > No changes are necessary to use HTTP Basic Auth - requests will > > > automatically use a .netrc file today. > > > > > > An optional dependency on kerberos and requests-kerberos can be added > to > > > support SPNEGO authentication. > > > Timeline > > > > > > I would like to land support for HTTP Basic Auth and SPNEGO in 0.8.0 > and > > > deprecate the SessionKey-based API for authentication in favor of fully > > > transport-based authentication. > > > > > > In 0.9.0 I propose removing SessionKey from the API entirely along with > > > SessionValidator from the scheduler. > > > > > > > > > > > -- > > Zameer Manji > > >
