Re: FW: Solr and LCF security at query time
Hi Karl, Apologies for the delayed reply. I've been away on business, and in the middle of a product release, so it's been a busy time... In response to your eariler questions: The 'AND/OR' filter query, will ultimately map down Lucene Boolean clauses, although the point at which these are done is slightly different. I think I am correct in my understanding that with filter queries, the results are filtered 'post-Lucene', but are separately (Solr) cached, so you get a hit on the first search, but then benefit from cached hits on subsequent searches. The lower-level 'MUST NOT/SHOULD' etc. clauses are applied at the Lucene query directly, so don't have separate Solr caching. I've not benchmarked the two, so one or other might be slower/faster for various search scenarios. In any case, I believe either technique can be employed in either 1834 or 1872. With regards schema extension, I believe we need to be very careful here, as requiring index-time storage of access control data will pose a problem for any use cases where the access control needs to change (maybe often, maybe only occasionally). I'm trying to think of a use case where this wouldn't at least potentially be the case, and I can't think of one, but perhaps I'm not truly understanding what exactly is stored in the __ALLOW_TOKEN__ and __DENY_TOKEN__ fields, and how/where subsequent acl changes would fit in (e.g. let's say someone has left my organization, do I have to update documents to remove his/her access?). Also, would such indexed tokens be entirely 'document-context-free'? I.e. Would the same type/format of tokens be used for data from different sources (e.g. NTFS files, network streams, NFS, web pages, etc.). Will the tokens be compatible with multiple and/or changing authorities (e.g. AD, documentum, LDAP, custom, etc.)? I like the idea of an LCF plugin to hold the acl data. I admit, I've not had enough time to look into how this might look at the moment, but it sounds like it could be a good way to hold generic (authority-agnostic) acl data, and [hopefully] not have to tie it to document data at index-time. I hope this makes sense, but if I've misunderstood the proposed mechanism, please correct me. Would the __ALLOW_TOKEN__ et al fields store, for example, SID information? Thanks, Peter On Tue, Apr 27, 2010 at 10:21 PM, karl.wri...@nokia.com wrote: Ok, not hearing back from Peter, I've done some Solr research and written some code that might work. The approach I've taken is most similar to SOLR 1834, other than the LCF-centric logic. Hopefully there will be a chance to try this out in a full end-to-end way on the weekend, after which I will submit it to the Solr team (where I think it most naturally would be built and delivered). What it's going to need is either a static or dynamic schema addition to define __ALLOW_TOKEN__document, __DENY_TOKEN__document, __ALLOW_TOKEN__share, and __DENY_TOKEN__share fields. These should be string, multivalued fields (I think). It would be great if these could be made a default part of Solr; similarly, it would be good if the new search component was predelivered with Solr and mentioned (even if commented out) in the example solrconfig.xml file. The only other thing that needs to be done to hook up the search component is to include a configuration parameter describing the base URL of the LCF authority service. Plus, as I said earlier, we still don't have a canned solution for authentication yet - although I feel that will be straightforward. Comments welcome... Karl From: Wright Karl (Nokia-S/Cambridge) Sent: Tuesday, April 27, 2010 8:20 AM To: connectors-dev@incubator.apache.org; d...@lucene.apache.org Cc: connectors-u...@incubator.apache.org; lucene-...@apache.org Subject: RE: FW: Solr and LCF security at query time Hi Peter, I finally had a moment to review the SOLR 1872 and SOLR 1834 contributions in detail, and have a couple of SOLR-related questions. Both contributions rely on a SearchComponent to work their magic. However, it also appears that each modifies the user query in a different way. 1834 uses MUST, MUST_NOT, and SHOULD filter items, while 1872 uses standard AND and OR filterquery clauses. Both of them are constructed using Solr FilterQuery objects. Here are my questions: (1) I am not conversant enough with Solr yet to know the difference between the different kinds of clause structure. Do you know if there is a difference? For example, is there any possibility that AND/OR clauses can permit documents to be seen that should not be seen? (MUST and MUST_NOT sound a lot more definite...) (2) Are Solr FilterQuery objects applied to constructing the query that will be sent to Lucene? Or are they applied by Solr after-the-fact to the resultset? Or, is it a combination of the two, depending on the details of your actual filter clause? I also haven't heard much from you
RE: FW: Solr and LCF security at query time
With regards schema extension, I believe we need to be very careful here, as requiring index-time storage of access control data will pose a problem for any use cases where the access control needs to change (maybe often, maybe only occasionally). I'm trying to think of a use case where this wouldn't at least potentially be the case, and I can't think of one, but perhaps I'm not truly understanding what exactly is stored in the __ALLOW_TOKEN__ and __DENY_TOKEN__ fields, and how/where subsequent acl changes would fit in (e.g. let's say someone has left my organization, do I have to update documents to remove his/her access?). Usually the way this works is that the user's account is locked out so they can't log in. The authority service picks up this change, and it therefore takes place immediately. Bear in mind that this particular model has been employed by MetaCarta for more than five years in the field with clients such as pretty near all the major oil companies, many U.S. government agencies, the U.S. military, etc. In that time we have not heard even one complaint about the security model. Karl From: ext Peter Sturge [mailto:peter.stu...@googlemail.com] Sent: Wednesday, April 28, 2010 7:18 AM To: d...@lucene.apache.org Cc: connectors-dev@incubator.apache.org; connectors-u...@incubator.apache.org; lucene-...@apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, Apologies for the delayed reply. I've been away on business, and in the middle of a product release, so it's been a busy time... In response to your eariler questions: The 'AND/OR' filter query, will ultimately map down Lucene Boolean clauses, although the point at which these are done is slightly different. I think I am correct in my understanding that with filter queries, the results are filtered 'post-Lucene', but are separately (Solr) cached, so you get a hit on the first search, but then benefit from cached hits on subsequent searches. The lower-level 'MUST NOT/SHOULD' etc. clauses are applied at the Lucene query directly, so don't have separate Solr caching. I've not benchmarked the two, so one or other might be slower/faster for various search scenarios. In any case, I believe either technique can be employed in either 1834 or 1872. With regards schema extension, I believe we need to be very careful here, as requiring index-time storage of access control data will pose a problem for any use cases where the access control needs to change (maybe often, maybe only occasionally). I'm trying to think of a use case where this wouldn't at least potentially be the case, and I can't think of one, but perhaps I'm not truly understanding what exactly is stored in the __ALLOW_TOKEN__ and __DENY_TOKEN__ fields, and how/where subsequent acl changes would fit in (e.g. let's say someone has left my organization, do I have to update documents to remove his/her access?). Also, would such indexed tokens be entirely 'document-context-free'? I.e. Would the same type/format of tokens be used for data from different sources (e.g. NTFS files, network streams, NFS, web pages, etc.). Will the tokens be compatible with multiple and/or changing authorities (e.g. AD, documentum, LDAP, custom, etc.)? I like the idea of an LCF plugin to hold the acl data. I admit, I've not had enough time to look into how this might look at the moment, but it sounds like it could be a good way to hold generic (authority-agnostic) acl data, and [hopefully] not have to tie it to document data at index-time. I hope this makes sense, but if I've misunderstood the proposed mechanism, please correct me. Would the __ALLOW_TOKEN__ et al fields store, for example, SID information? Thanks, Peter On Tue, Apr 27, 2010 at 10:21 PM, karl.wri...@nokia.commailto:karl.wri...@nokia.com wrote: Ok, not hearing back from Peter, I've done some Solr research and written some code that might work. The approach I've taken is most similar to SOLR 1834, other than the LCF-centric logic. Hopefully there will be a chance to try this out in a full end-to-end way on the weekend, after which I will submit it to the Solr team (where I think it most naturally would be built and delivered). What it's going to need is either a static or dynamic schema addition to define __ALLOW_TOKEN__document, __DENY_TOKEN__document, __ALLOW_TOKEN__share, and __DENY_TOKEN__share fields. These should be string, multivalued fields (I think). It would be great if these could be made a default part of Solr; similarly, it would be good if the new search component was predelivered with Solr and mentioned (even if commented out) in the example solrconfig.xml file. The only other thing that needs to be done to hook up the search component is to include a configuration parameter describing the base URL of the LCF authority service. Plus, as I said earlier, we still don't have a canned solution
Re: FW: Solr and LCF security at query time
Hi Karl, Yes, I don't doubt that using an external mechanism such as AD lockout will work for those and other environments. I guess it comes down to the difference between bespoke consultancy-type solutions and general-purpose product solutions, of which the requirements are often very different. For a general Access Control solution integrated into Solr, assumptions on the presence/type of such external controls can't, and should't be assumed. If they are/must be assumed, one of the core reasons for adding the new functionality is missing. As a starting point, for a general purpose access control system, at least the following questions need to be addressed: * What happens when access control needs to change? * What happens if access control needs to change often (e.g. more than several times a day)? * Can the access control cope with multiple data source types, without the need for custom code, including data with no attached acl information? * If I change my access control, how is 'offline' data affected? (e.g. backed-up data) * Will the access control satisfy regulatory compliance specs on it own, or is an external mechanism required? (currently, Solr requires an external mechanism, but so also the proposed solution) As you might have guessed, I've been down this road before, and the productization of security control has many facets, and these, as a general rule, need to be addressed differently in products than in site-specific deployments - mainly because products can't assume the envinroment(s) they will run in (e.g. Active Directory). The good thing is, there is a good alternative - that is: to store access control information separately from indexed data and separately from an authority. To me, that's where the beauty of an LCF plugin architecture lives. Then, the task is to provide the integration tools (and it sounds like LCF is very well suited for this) to deliver the 'bridge' between content and authorization. (as you quite rightly said, authentication is a separate, albeit related, subject) Thanks, Peter On Wed, Apr 28, 2010 at 12:46 PM, karl.wri...@nokia.com wrote: With regards schema extension, I believe we need to be very careful here, as requiring index-time storage of access control data will pose a problem for any use cases where the access control needs to change (maybe often, maybe only occasionally). I'm trying to think of a use case where this wouldn't at least potentially be the case, and I can't think of one, but perhaps I'm not truly understanding what exactly is stored in the __ALLOW_TOKEN__ and __DENY_TOKEN__ fields, and how/where subsequent acl changes would fit in (e.g. let's say someone has left my organization, do I have to update documents to remove his/her access?). Usually the way this works is that the user's account is locked out so they can't log in. The authority service picks up this change, and it therefore takes place immediately. Bear in mind that this particular model has been employed by MetaCarta for more than five years in the field with clients such as pretty near all the major oil companies, many U.S. government agencies, the U.S. military, etc. In that time we have not heard even one complaint about the security model. Karl -- *From:* ext Peter Sturge [mailto:peter.stu...@googlemail.com] *Sent:* Wednesday, April 28, 2010 7:18 AM *To:* d...@lucene.apache.org *Cc:* connectors-dev@incubator.apache.org; connectors-u...@incubator.apache.org; lucene-...@apache.org *Subject:* Re: FW: Solr and LCF security at query time Hi Karl, Apologies for the delayed reply. I've been away on business, and in the middle of a product release, so it's been a busy time... In response to your eariler questions: The 'AND/OR' filter query, will ultimately map down Lucene Boolean clauses, although the point at which these are done is slightly different. I think I am correct in my understanding that with filter queries, the results are filtered 'post-Lucene', but are separately (Solr) cached, so you get a hit on the first search, but then benefit from cached hits on subsequent searches. The lower-level 'MUST NOT/SHOULD' etc. clauses are applied at the Lucene query directly, so don't have separate Solr caching. I've not benchmarked the two, so one or other might be slower/faster for various search scenarios. In any case, I believe either technique can be employed in either 1834 or 1872. With regards schema extension, I believe we need to be very careful here, as requiring index-time storage of access control data will pose a problem for any use cases where the access control needs to change (maybe often, maybe only occasionally). I'm trying to think of a use case where this wouldn't at least potentially be the case, and I can't think of one, but perhaps I'm not truly understanding what exactly is stored in the __ALLOW_TOKEN__ and __DENY_TOKEN__ fields, and how
Re: FW: Solr and LCF security at query time
Hi Karl, I wasn't trying to to put pay to your design proposal, really the opposite - to highlight requirements that have found to be necessary for customers/users, and to hopefully get the best functionality for the product. If you feel I've put you out on any of the issues raised, then I apologize for that, it was certainly not my intention. Peter
RE: FW: Solr and LCF security at query time
Ok, not hearing back from Peter, I've done some Solr research and written some code that might work. The approach I've taken is most similar to SOLR 1834, other than the LCF-centric logic. Hopefully there will be a chance to try this out in a full end-to-end way on the weekend, after which I will submit it to the Solr team (where I think it most naturally would be built and delivered). What it's going to need is either a static or dynamic schema addition to define __ALLOW_TOKEN__document, __DENY_TOKEN__document, __ALLOW_TOKEN__share, and __DENY_TOKEN__share fields. These should be string, multivalued fields (I think). It would be great if these could be made a default part of Solr; similarly, it would be good if the new search component was predelivered with Solr and mentioned (even if commented out) in the example solrconfig.xml file. The only other thing that needs to be done to hook up the search component is to include a configuration parameter describing the base URL of the LCF authority service. Plus, as I said earlier, we still don't have a canned solution for authentication yet - although I feel that will be straightforward. Comments welcome... Karl From: Wright Karl (Nokia-S/Cambridge) Sent: Tuesday, April 27, 2010 8:20 AM To: connectors-dev@incubator.apache.org; d...@lucene.apache.org Cc: connectors-u...@incubator.apache.org; lucene-...@apache.org Subject: RE: FW: Solr and LCF security at query time Hi Peter, I finally had a moment to review the SOLR 1872 and SOLR 1834 contributions in detail, and have a couple of SOLR-related questions. Both contributions rely on a SearchComponent to work their magic. However, it also appears that each modifies the user query in a different way. 1834 uses MUST, MUST_NOT, and SHOULD filter items, while 1872 uses standard AND and OR filterquery clauses. Both of them are constructed using Solr FilterQuery objects. Here are my questions: (1) I am not conversant enough with Solr yet to know the difference between the different kinds of clause structure. Do you know if there is a difference? For example, is there any possibility that AND/OR clauses can permit documents to be seen that should not be seen? (MUST and MUST_NOT sound a lot more definite...) (2) Are Solr FilterQuery objects applied to constructing the query that will be sent to Lucene? Or are they applied by Solr after-the-fact to the resultset? Or, is it a combination of the two, depending on the details of your actual filter clause? I also haven't heard much from you in the last week or so - have you thought further about what you intend to do, and can you let me know whether you are still interested in developing an LCF plugin for Solr? Thanks, Karl -Original Message- From: ext Peter Sturge [mailto:peter.stu...@googlemail.com] Sent: Thursday, April 22, 2010 12:23 PM To: d...@lucene.apache.org Cc: connectors-dev@incubator.apache.org; connectors-u...@incubator.apache.org; lucene-...@apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, See inline... On Thu, Apr 22, 2010 at 4:57 PM, karl.wri...@nokia.com wrote: Hi Peter, The authority connectors don't perform authentication at this time. In fact, LCF has nothing to do with authentication at all - just authorization. The reason for this is because it is almost never the case that somebody wants to provide multiple credentials in order to be able see their results. Most enterprises who have multiple repositories authenticate against AD and then map AD user names to repository user names in order to access those repositories. If you noted my earlier posts from this morning, you may have noted that I'm looking at recommending JAAS plus sun's kerb5 login module for handling the authenticate against AD case, which would cover some 95%+ of the real world authentication needed out there. I did read your earlier post regarding this, and I totally agree with you - this is best handled 'upstream'. In fact, I use a JAAS plugin in other places in the product (not Solr) for authentication. Yes, the idea is to store SIDs in solr at index time. I don't know enough about solr to know what kinds of issues this might entail, but Lucene certainly has a model of metadata that's pretty flexible, so I don't think this would be difficult at all. Eric Hatcher also seemed to confirm my suspicions that this would not be a problem. It's certainly not a problem to store this data in Solr. The problem is more that you don't really *want* to store this data at index time. There are lots of reasons for not wanting to 'hard-code' SID data with documents in the index. Here's just a few: * What happens if/when you want to add explicit user access to some [group of] documents ? (i.e. not via a group) * What happens if you need to revoke or change a user's or group's access? * It's difficult to move/replicate the index
RE: FW: Solr and LCF security at query time
Looking around for no-Apache java-only solutions to the AD authentication problem, it seems to me that what we mainly have available is JAAS plus the following JAAS login module: com.sun.security.auth.module.Krb5LoginModule ... which should permit AD authentication to take place, if properly configured. So, we *could* stipulate that the search component receive credentials, somehow, upon being called, and then authenticate each time. (There's a ticket cache involved, so this is not as insane as it sounds). But this architecture option makes me twitchy because I am unclear how exactly this would help Tomcat interact with the browser to manage signon for a web application. So it might be better to push the authentication itself upstream into a module meant to be plugged into Tomcat, and have Solr just receive and deal with the resulting ticket, and/or an authenticated, domain-qualified user name. The task of the LCF Solr search component or filter would then be to do the following: (1) Get hold of the ticket/authenticated user name, which will probably come in as some attribute to the search that's presented to Solr. (Someone needs to specify what this attribute is called still). (2) Invoke a configured LCF authority service with that user name, via http, and get back a list of access tokens for the user (3) Form the search expression with the user's access tokens (if it's a search component), or filter the results using those access tokens (if it's a filter), remembering that every document that's participating in security should have __ACCESS_TOKEN__document and __DENY_TOKEN__document metadata I've also been pondering whether which we should build: a search component or filter? I think there are advantages to both, so I think we should build both, and let people use what they need. I think the technical aspects of building the Solr component are well understood by this group, so the only open issue remains how to build a JAAS-based AD authentication module for tomcat that would do what we needed. I'll be doing more research as time permits... Karl From: Wright Karl (Nokia-S/Cambridge) Sent: Wednesday, April 21, 2010 8:02 PM To: connectors-u...@incubator.apache.org; lucene-...@apache.org; connectors-dev@incubator.apache.org Subject: RE: FW: Solr and LCF security at query time Hi Peter, I just committed the promised changes to the LCF Solr output connector. ACL metadata will now be posted to the Solr Http interface along with the document as the two following fields: __ACCESS_TOKEN__document __DENY_TOKEN__document There will, of course, potentially be multiple values for each of these two fields. Hope this helps, Karl From: ext Peter Sturge [mailto:peter.stu...@googlemail.com] Sent: Tuesday, April 20, 2010 6:51 PM To: connectors-u...@incubator.apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, Thanks for the info. I'll have a look at the link and try to take in as much sugar as my insulin levels will handle... It sounds like the necessary interface(s) are already in LCF - just a matter of implementing them in the Solr 1872 plugin. I'll need to digest the LCF stuff to get to grips with it..please bear with me while I do that... When you say: The LCF solr output connection doesn't yet do this, but it is trivial for me to make that happen. Do you mean a mechanism by which solr.war can get url et al info from its parent container (Tomcat, Jetty etc.), or have I misinterpreted this? Thanks, Peter On Tue, Apr 20, 2010 at 11:05 PM, karl.wri...@nokia.commailto:karl.wri...@nokia.com wrote: Hi Peter, I'm the principal committer for LCF, but I don't know as much about Solr as I ought to, so it sounds like a potentially productive collaboration. LCF does exactly what you are looking for - the only issue at all is that you need to fetch a URL from a webapp to get what you are looking for. The plugs are all inside LCF for different kinds of repositories. Here's a link that might help with drinking the LCF koolaid, as it were: https://cwiki.apache.org/confluence/display/CONNECTORS/Lucene+Connectors+Framework+concepts The url would be something like this (on a locally installed tomcat-based LCF instance): http://localhost:8080/lcf-authority-service/useracls?username=someusern...@somedomain.com ... and this fetch returns something like: TOKEN:xxx TOKEN:yyy TOKEN:zzz ... which represent the amalgamated tokens for all of the defined authorities, and by some strange coincidence ( ;-) ) are compatible with certain pieces of metadata that have been passed into Solr with each document - one set of Allow tokens, and a second set of Deny tokens. The LCF solr output connection doesn't yet do this, but it is trivial for me to make that happen. Does this sound plausible to you? Karl From: ext Peter
RE: FW: Solr and LCF security at query time
Hi Peter, I've attached a diagram that is not in the wiki as of yet, and I'll try to answer your questions. Are the ACCESS_TOKEN and DENY_TOKEN values whatever have been stored for a particular user in the underlying acl store (e.g. Active Directory)? How does AD and/or LCF handle storing such data in its schema? (does AD needs its schema extended?) Presumably, any such AD fields would need to be queried for effective rights in order to cater for group membership allows and denies. The ACCESS_TOKEN and DENY_TOKEN values are, in one sense, arbitrary strings that represent a contract between an LCF authority connection and the LCF repository connection that picks up the documents (from wherever). These tokens thus have no real meaning outside of LCF. You must regard them as opaque. The contract, however, states that if you use the LCF authority service to obtain tokens for an authenticated user, you will get back a set that is CONSISTENT with the tokens that were attached to the documents LCF sent to Solr for indexing in the first place. So, you don't have to worry about it, and that's kind of the idea. So you imagine the following flow: (1) Use LCF to fetch documents and send them to Solr (2) When searching, use the LCF authority service to get the desired user's access tokens (3) Either filter the results, or modify the query, to be sure the access tokens all match up properly For the AD authority, the LCF access tokens consist, in part, of the user's SIDs. For other authorities, the access tokens are wildly different. You really don't want to know what's in them, since that's the job of the LCF authority to determine. ;-) LCF is not, by the way, joined at the hip with AD. However, in practice, most enterprises in the world use some form of AD single signon for their web applications, and even if they're using some repository with its own idea of security, there's a mapping between the AD users and the repository's users. Doing that mapping is also the job of the LCF authority for that repository. Hope this helps. Also, I'm not expecting time miracles here, so don't sweat the schedule. Karl From: ext Peter Sturge [peter.stu...@googlemail.com] Sent: Thursday, April 22, 2010 4:27 AM To: d...@lucene.apache.org Cc: connectors-u...@incubator.apache.org; lucene-...@apache.org; connectors-dev@incubator.apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, Thanks for the quick turnaround. I'm in the middle of a product release for us, so I fear I won't be as quick as you... :-) I couldn't find a simple flow diagram or similar for LCF with regards security (probably looking in the wrong place). Perhaps you could help on these questions...? In SOLR-1872, the allows and denies are stored (in acl.xml) as sub-queries, which are then used as filter queries in a user's search. Are the ACCESS_TOKEN and DENY_TOKEN values whatever have been stored for a particular user in the underlying acl store (e.g. Active Directory)? How does AD and/or LCF handle storing such data in its schema? (does AD needs its schema extended?) Presumably, any such AD fields would need to be queried for effective rights in order to cater for group membership allows and denies. I guess I'm just trying to understand the architectural flow/storage/retrieval of data in the various parts of the system, but I admit, I need to do more research on this. After our product release, when I get a few more spare cycles, I can look at it in more detail. Many thanks! Peter On Thu, Apr 22, 2010 at 1:02 AM, karl.wri...@nokia.commailto:karl.wri...@nokia.com wrote: Hi Peter, I just committed the promised changes to the LCF Solr output connector. ACL metadata will now be posted to the Solr Http interface along with the document as the two following fields: __ACCESS_TOKEN__document __DENY_TOKEN__document There will, of course, potentially be multiple values for each of these two fields. Hope this helps, Karl From: ext Peter Sturge [mailto:peter.stu...@googlemail.commailto:peter.stu...@googlemail.com] Sent: Tuesday, April 20, 2010 6:51 PM To: connectors-u...@incubator.apache.orgmailto:connectors-u...@incubator.apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, Thanks for the info. I'll have a look at the link and try to take in as much sugar as my insulin levels will handle... It sounds like the necessary interface(s) are already in LCF - just a matter of implementing them in the Solr 1872 plugin. I'll need to digest the LCF stuff to get to grips with it..please bear with me while I do that... When you say: The LCF solr output connection doesn't yet do this, but it is trivial for me to make that happen. Do you mean a mechanism by which solr.war can get url et al info from its parent container (Tomcat, Jetty etc.), or have I misinterpreted this? Thanks, Peter On Tue
Re: FW: Solr and LCF security at query time
Hi Karl, Thanks very much for the diagram - Sorry about all the questions, but this raises a few new ones... What is the relationship between stored data (documents) and authorities' access/deny attributes? (do you have any examples of what an access_token value might contain?) One of the key requirements I've worked to adhere to in SOLR-1872 is to ensure there are no security or other dependencies of indexed data with any external repository - most notably the file system. There are many reasons for wanting this, but one of the main ones is that Solr-stored data is not always based on file data (or accessible file data). In fact, in my particular case, almost none of the indexed data comes from files. This is one reason why SOLR-1872 uses filter queries for its access/deny tokens - so that all the required information for access control completely resides within the Solr index itself. Is the LCF architecture acl 'mapping' between Solr fields (queries) and users, some external 'repository' (files) and users, or arbitrary data (e.g. either of these)? I hope that makes sense... Thanks! Peter On Thu, Apr 22, 2010 at 10:25 AM, karl.wri...@nokia.com wrote: Hi Peter, I've attached a diagram that is not in the wiki as of yet, and I'll try to answer your questions. Are the ACCESS_TOKEN and DENY_TOKEN values whatever have been stored for a particular user in the underlying acl store (e.g. Active Directory)? How does AD and/or LCF handle storing such data in its schema? (does AD needs its schema extended?) Presumably, any such AD fields would need to be queried for effective rights in order to cater for group membership allows and denies. The ACCESS_TOKEN and DENY_TOKEN values are, in one sense, arbitrary strings that represent a contract between an LCF authority connection and the LCF repository connection that picks up the documents (from wherever). These tokens thus have no real meaning outside of LCF. You must regard them as opaque. The contract, however, states that if you use the LCF authority service to obtain tokens for an authenticated user, you will get back a set that is CONSISTENT with the tokens that were attached to the documents LCF sent to Solr for indexing in the first place. So, you don't have to worry about it, and that's kind of the idea. So you imagine the following flow: (1) Use LCF to fetch documents and send them to Solr (2) When searching, use the LCF authority service to get the desired user's access tokens (3) Either filter the results, or modify the query, to be sure the access tokens all match up properly For the AD authority, the LCF access tokens consist, in part, of the user's SIDs. For other authorities, the access tokens are wildly different. You really don't want to know what's in them, since that's the job of the LCF authority to determine. ;-) LCF is not, by the way, joined at the hip with AD. However, in practice, most enterprises in the world use some form of AD single signon for their web applications, and even if they're using some repository with its own idea of security, there's a mapping between the AD users and the repository's users. Doing that mapping is also the job of the LCF authority for that repository. Hope this helps. Also, I'm not expecting time miracles here, so don't sweat the schedule. Karl From: ext Peter Sturge [peter.stu...@googlemail.com] Sent: Thursday, April 22, 2010 4:27 AM To: d...@lucene.apache.org Cc: connectors-u...@incubator.apache.org; lucene-...@apache.org; connectors-dev@incubator.apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, Thanks for the quick turnaround. I'm in the middle of a product release for us, so I fear I won't be as quick as you... :-) I couldn't find a simple flow diagram or similar for LCF with regards security (probably looking in the wrong place). Perhaps you could help on these questions...? In SOLR-1872, the allows and denies are stored (in acl.xml) as sub-queries, which are then used as filter queries in a user's search. Are the ACCESS_TOKEN and DENY_TOKEN values whatever have been stored for a particular user in the underlying acl store (e.g. Active Directory)? How does AD and/or LCF handle storing such data in its schema? (does AD needs its schema extended?) Presumably, any such AD fields would need to be queried for effective rights in order to cater for group membership allows and denies. I guess I'm just trying to understand the architectural flow/storage/retrieval of data in the various parts of the system, but I admit, I need to do more research on this. After our product release, when I get a few more spare cycles, I can look at it in more detail. Many thanks! Peter On Thu, Apr 22, 2010 at 1:02 AM, karl.wri...@nokia.commailto: karl.wri...@nokia.com wrote: Hi Peter, I just committed the promised changes to the LCF Solr output connector
RE: FW: Solr and LCF security at query time
Hi Peter, * What happens if/when you want to add explicit user access to some [group of] documents ? (i.e. not via a group) In LCF, you change the permissions on the appropriate resource, and then you run your LCF job again to update those permissions. Since LCF is an incremental crawler, it is smart enough to only reindex those documents whose permissions have changed, which makes it a fairly fast operation on most repositories. Also, in my experience at MetaCarta, this is a relatively infrequent kind of situation, and most enterprises are pretty resilient against there being a reasonable delay in getting document permissions updated in an index. However, if this is a concern in your environment, your main alternative is to go directly to the repository on every document as you filter a resultset. That's slow for most situations, perhaps not for a local acl.xml file. Performance might be improved with caching, but only if you knew that the same results would be returned for multiple queries. * What happens if you need to revoke or change a user's or group's access? I presume you mean a user/group's access to specific documents - which has the same answer as above. If you actually mean the more typical case, where a user is locked out, or loses/gains group access, that of course happens at authority time, so it is instantaneous. * It's difficult to move/replicate the index to another domain Sure. If this is something you intend on doing a lot, this is not a solution that will work for you. I don't think we ever had any of MetaCarta's clients even *think* of doing this, however. ;-) Probably because lots of other stuff breaks as well. * For AD, SIDs are generally not meant to be stored long term outside of AD, as they can be changed (this doesn't happen often, but it can happen after an AD rebuild, domain type upgrade, data recovery etc.) Any infrequent operation is not much of a concern to me, since LCF keeps track of any changes and will pick them up on the next crawl (and do the minimum possible to update the index, as well). Thanks, Karl -Original Message- From: ext Peter Sturge [mailto:peter.stu...@googlemail.com] Sent: Thursday, April 22, 2010 12:23 PM To: d...@lucene.apache.org Cc: connectors-dev@incubator.apache.org; connectors-u...@incubator.apache.org; lucene-...@apache.org Subject: Re: FW: Solr and LCF security at query time Hi Karl, See inline... On Thu, Apr 22, 2010 at 4:57 PM, karl.wri...@nokia.com wrote: Hi Peter, The authority connectors don't perform authentication at this time. In fact, LCF has nothing to do with authentication at all - just authorization. The reason for this is because it is almost never the case that somebody wants to provide multiple credentials in order to be able see their results. Most enterprises who have multiple repositories authenticate against AD and then map AD user names to repository user names in order to access those repositories. If you noted my earlier posts from this morning, you may have noted that I'm looking at recommending JAAS plus sun's kerb5 login module for handling the authenticate against AD case, which would cover some 95%+ of the real world authentication needed out there. I did read your earlier post regarding this, and I totally agree with you - this is best handled 'upstream'. In fact, I use a JAAS plugin in other places in the product (not Solr) for authentication. Yes, the idea is to store SIDs in solr at index time. I don't know enough about solr to know what kinds of issues this might entail, but Lucene certainly has a model of metadata that's pretty flexible, so I don't think this would be difficult at all. Eric Hatcher also seemed to confirm my suspicions that this would not be a problem. It's certainly not a problem to store this data in Solr. The problem is more that you don't really *want* to store this data at index time. There are lots of reasons for not wanting to 'hard-code' SID data with documents in the index. Here's just a few: * What happens if/when you want to add explicit user access to some [group of] documents ? (i.e. not via a group) * What happens if you need to revoke or change a user's or group's access? * It's difficult to move/replicate the index to another domain * For AD, SIDs are generally not meant to be stored long term outside of AD, as they can be changed (this doesn't happen often, but it can happen after an AD rebuild, domain type upgrade, data recovery etc.) These and other senarios mean re-indexing the stored data. When the index is huge, this is non-trivial (time-wise). There are not uncommon scenarios where user/group access control can change multiple times in one day. There might be a way of storing acl data in a payload or similar, but I'm not sure how that would work across millions of [arbitrarily grouped ] documents (I'm not familiar enough with payloads to know