Hiya Geoff, You’re right, there is a non-trivial overhead in calculating view responses that need to pull from every shard. On the other hand, maintaining a unique database file for every user is quite problematic at scale. It works OK up to several thousand users, but eventually you start running into a lot of operational headaches. Running a million databases in a cluster is possible but painful.
I have some detailed thoughts about how we can improve the efficiency of queries scoped to a single user in a large sharded database, but that’s a topic for another thread :) Jan - wow, look at that! I’ll take a close look over the next couple of hours but a quick scan is encouraging. Adam > On Nov 15, 2017, at 5:30 PM, Geoffrey Cox <redge...@gmail.com> wrote: > > Hey Jan, > > I've been trying to solve a similar problem from a different angle using > efficient and scalable replication via spiegel > <https://github.com/redgeoff/spiegel>. I'm super excited that you are > drafting this level of access, but my major concern is on performance. From > what I gather, if you combine all the db-per-user docs into a single DB > then you'll have a massive DB. I know CouchDB is good at sharding, but > isn't there a significant performance implication when a user's docs are > being pulled from multiple shards on different servers? What about the > added overhead of calculating cross-server views, etc... > > When I think about how big companies, e.g. Facebook, solve these types of > problems, I imagine that they create a denormalized DB per user. Among > other things, this design allows the set of data that a user needs to be > relatively small and live on less servers per user. Doesn't this lead to > better performance? > > Even if this new level of access doesn't solve the db-per-user case > entirely, it will still be a useful addition as it would allow for more > data to be shared and less of a create a DB-per-role setup. So, I'm all for > it! > > I'll take a closer look at these notes when I have some time, but I just > wanted to get you my high-level thoughts now. I'm sorry if any of this has > been based on some wild assumptions :) > > Exciting stuff! > > Geoff > > On Wed, Nov 15, 2017 at 1:35 PM Jan Lehnardt <j...@apache.org> wrote: > >> Hi all, >> >> in the midst of handling the security stuff I had a moment of clarity how >> the often requested per document permissions could be implemented. We had >> discussed a potential approach extensively in the February Boston Developer >> Summit (notes here: >> https://lists.apache.org/thread.html/09a5686bca8049010b82796cc0fe99ef27aed4983a3f02fd6956259f@%3Cdev.couchdb.apache.org%3E >> ) >> >> What was so alluring about this proposal was that it solves per doc access >> control and per-user-db in one go. E.g. it would be able to share a single >> database with multiple distrusting users, allow them to have their own set >> of views, and even independently use their share of a single database as a >> replication endpoint without interfering with any of the other users on >> that database. >> >> I gave it a shot. Essentially, we need to build new indexes: by-access-id >> and by-access-seq to make all that work. I’m just focussing on the core of >> this, trying to re-use the existing couch_mrview/couch_index machinery as >> much as possible. Strictly, for replication only by-access-seq would be >> required, but by-update-id is a little easier to do, so I’ve done that >> first, and I believe the results are encouraging. >> >> I’ve put a diff against master into a gist for your perusal: >> >> https://gist.github.com/janl/20b218a3f0eafbf963ee28780261f9fc >> >> >> The core bits are: >> >> >> https://gist.github.com/janl/20b218a3f0eafbf963ee28780261f9fc#file-by-access-id-diff-L189-L215 >> >> and >> >> >> https://gist.github.com/janl/20b218a3f0eafbf963ee28780261f9fc#file-by-access-id-diff-L189-L215 >> >> Here’s an example Doc: >> >> { >> "_id":"1fb94bf8c3d5a73745f3cc4f5f000a8d”, >> "_rev":"4-bcbc975e61bdb80f3de1b87f6cad6a76”, >> "_access":["b”] >> } >> >> It shows up for user b: >> >> >> curl b:b@127.0.0.1:15984/a/_all_docs >> >> {"total_rows”:2,"offset":0,"rows":[ >> >> {"id":"1fb94bf8c3d5a73745f3cc4f5f000a8d","key":["b","1fb94bf8c3d5a73745f3cc4f5f000a8d"],"value":"4-bcbc975e61bdb80f3de1b87f6cad6a76”} >> ]} >> >> But not for user c: >> >> >>> curl c:c@127.0.0.1:15984/a/_all_docs >> >> {"total_rows”:2,"offset":2,"rows":[ >> >> ]} >> >> >> * * * >> >> >> I’d like to get some general design feedback on this approach to find out >> if it is worth pursuing further. See “Next Steps” way below for my thinking >> on how to get by-access-seq going. >> >> The rest of this email are my notes from reading the source and trying to >> explain my thinking as well as guide folks that might not be very familiar >> with the CouchDB sources to follow along what is happening. >> >> I’d especially like to get some feedback about this from some of the folks >> here who don’t spend their days in the main Erlang codebase :) >> >> Let me know what you think. >> >> Thanks! >> Jan >> >> * * * >> >> CouchDB Access Notes >> >> Background: >> https://lists.apache.org/thread.html/09a5686bca8049010b82796cc0fe99ef27aed4983a3f02fd6956259f@%3Cdev.couchdb.apache.org%3E >> >> # Overview >> >> To solve the problems with the db-per-user pattern, we want to introduce >> document level access control. The result should be a single CouchDB >> database that can be used by multiple mutually untrusting users while >> retaining CouchDB’s full semantics. >> >> // TODO: link to appendix: problems with db-per-user >> >> We decided on an approach to define access control in documents with a new >> property `_access` which is specified as an array of strings and arrays. >> Strings represent usernames and roles, sub-arrays are used as logical AND, >> elements in the top level array are used as logical OR. For example. an >> _access field with the value [[‘management’, ‘senior’], ‘ceo-jane’] would >> allow everyone with the roles ‘management’ AND ‘senior’, OR the user >> ‘ceo-jane’ access to that doc. but not e.g. users with roles ‘development’, >> ‘senior’, nor user ‘vp-jenn’. >> >> To achieve main CouchDB semantics, we need to introduce new behaviour for >> the _all_docs and _changes endpoints. The plan is to special case-this >> based on the authenticated user context (userCtx, e.g, username and >> associated roles, after authentication). >> >> The existing by-id and by-seq indexes are not equipped to efficiently >> return results per user, so we are introducing two new indexes (either can >> be optionally configured, depending on the use-case and performance and >> storage needs): by-access-id and by-access-seq. In contrast with by-id and >> by-seq, these indexes are not stored in the main database file, but in a >> separate file, ideally managed by the existing couch_index infrastructure. >> >> >> # Development considerations >> >> This first spike is only concerned with getting per-access-id to work with >> minimal effort. >> >> To get started, let’s look at how _all_docs works today using the by-id >> index. >> >> ## The Anatomy of a Clustered _all_docs Request >> >> CouchDB’s clustering layer consists of three main modules: chttpd, fabric >> and refi. chttpd’s job is to handle everything HTTP and route requests to >> the right place in the rest of the code. It’s a HTTP router, mapping URLs, >> request methods and options to handler functions that do with the work the >> requests are specified to fulfil. >> >> fabric’s job is to distribute a single request from the outside to >> multiple nodes of the cluster. Some requests require only talking to the >> local node, but that’s less important for the moment. fabric includes >> fabric_rpc, a module that turns a request to the cluster into one or more >> requests to other nodes in the cluster. >> >> rexi’s job is know about the cluster state: which nodes are in the >> cluster, which of them are active/reachable/failed, which shards live on >> which nodes. fabric uses rexi to know which nodes to contact for which >> shards. >> >> After a bit of indirection, we find ourselves at the first >> _all_docs-specific function in chttpd_db.erl: all_docs_view/4: >> >> ``` >> all_docs_view(Req, Db, Keys, OP) -> >> Args0 = couch_mrview_http:parse_params(Req, Keys), >> Args1 = Args0#mrargs{view_type=map}, >> Args2 = couch_mrview_util:validate_args(Args1), >> Args3 = set_namespace(OP, Args2), >> Options = [{user_ctx, Req#httpd.user_ctx}], >> Max = chttpd:chunked_response_buffer_size(), >> VAcc = #vacc{db=Db, req=Req, threshold=Max}, >> {ok, Resp} = fabric:all_docs(Db, Options, fun >> couch_mrview_http:view_cb/2, VAcc, Args3), >> {ok, Resp#vacc.resp}. >> ``` >> >> The first five lines handle query options and request parameters or >> arguments. The next three lines are the bulk of the job: start a response, >> call fabric:all_docs/5 with a callback to handle rows. The last line >> returns the accumulator that is returned by fabric:all_docs/5. >> >> fabric:all_docs/5 is a thin wrapper around fabric_view_all_docs:go/5. >> Before we jump down, we notice that there is also a >> fabric_view_changes.erl, which we should remember for the next iteration >> when we implement by-access-seq. >> >> go/5 comes in two variants and we’ll ignore the second here for the >> moment, because it is a performance optimisation. The main work for go/5 is >> in the top third of the function. First it gets all shards for the current >> database from mem3, then it starts a fabric_rpc worker for each shard, and >> then waits for the results to come back by calling go/6 with all workers. >> The bottom two thirds are timeout and error handling. >> >> go/6 registers the handle_message/3 function as the callback for >> rexi_utils’ recv/6 (read “receive”) function. >> >> handle_message/3 comes in a number of variants to handle rexi errors, >> receiving metadata, receiving result rows and a notification “complete” >> about all rows having been sent. >> >> Our next level down is looking into fabric_rpc and how it handles all_docs >> requests. fabric_rpc/3 is again a short wrapper, this time around >> couch_mrview:query_all_docs/4 which is the node-local function that handles >> querying. >> >> couch_mrview includes a bunch of functions map/reduce views. It seems like >> a natural place doing our distinction between a normal by-id request and a >> by-access-id request. >> >> I’m skipping a step here, but with a little printf-debugging, I’ve found >> out that the `Db` variable we get passed in, includes the authenticated >> userCtx including username and any roles. We can use couch_db:is_admin/1 >> to get a boolean back for the distinction we are going to have to make: >> >> ``` >> query_all_docs(Db, Args0, Callback, Acc) -> >> case couch_db:is_admin(Db) of >> true -> query_all_docs_admin(Db, Args0, Callback, Acc); >> false -> query_all_docs_access(Db, Args0, Callback, Acc) >> end. >> ``` >> >> query_all_docs_admin/4 is the existing query_all_docs/4 function and we’re >> introducing query_all_docs_access/4, that we now have to fill out with >> querying our view. >> >> Before we can do that, we need to understand how view work. >> >> ## The Anatomy of a View Request >> >> Querying a view has three stages: >> >> 1. define the view >> 2. build the view index >> 3. query the view index >> >> A view definition is always in a design document. It can be one or >> JavaScript map/reduce functions, Erlang map/reduce functions, or a mango >> index definition. >> >> // TODO: link all these view definition options. >> >> Building the view index is an implicit step in CouchDB. View indexes are >> refreshed at query time, but only if there were any changes in the database >> since the last query. If no refresh is needed, the view result is returned >> from the index directly. >> >> // TODO: explain query_server >> >> Querying indexes follows a similar path through chttpd, fabric, rexi, >> fabric_rpc down to the per-node handlers in couch_mrview. Just a few lines >> below couch_mrview:query_all_docs/4 we find query_view/5 which decides >> between map and reduce requests. We care about map-only for now. >> query_view/5 is preceded by query_view/6 which includes a call to >> couch_mrview_util:get_view/4 which looks like it is where we want to look >> next, as the map_fold/5 called by query_view/5 is about looping over rows. >> We hope we can re-use all that logic, and maybe get_view/4 lets us find out >> how we can have it return our new view. >> >> get_view/4 calls get_view_index_state/4 which in turn calls >> get_view_index_pid/4 that finally calls into couch_index_server:get_index/4 >> which looks like it returns the index for our request. Let’s have a look. >> >> get_index/4 will dive into get_index/2 eventually and that looks indeed >> like where we need to look. In there, we look up view index in an ETS table >> (an in-memory database), and if it can’t find it there, start a new one. >> Either way, a view index is returned. The lookup is by DbName and >> Sig(nature), an md5 hash over the `views` property in a design doc, that >> also corresponds to the *.view filename of the view index. >> >> >> ## Faking the index >> >> So how would we get this to return the index we want to query? We need to >> create an index definition that matches the design doc `views` hash. Hm. >> >> It is relatively easy to produce a map function that behaves like we want: >> >> function (doc) { >> var _access = doc.access >> if (!_access) { return } >> if (!isArray(_access) || _access,length === 0) { return } >> _access.forEach( function (user_or_role) { >> emit([user_or_role, doc._id], doc._rev) >> }) >> } >> >> At query time, we’d have to match the requesting username and roles >> against the first element in the key-array and return the results, while >> replacing the key-array with the second element (the doc _id). All this >> doesn’t sound too hard. Good. >> >> One snag though: if we think ahead and try to see how we could implement >> by-access-changes we get stuck: a view does not include rows for deleted >> documents while _changes does. In addition, the update sequence for a >> document is not available in a map function. So a regular view can not be >> used here. >> >> The filtering of deleted docs from a view index happens in >> couch_mrview:map_fold/3. So if we could augment that for our internal view >> requests, that could get us a long way towards reusing the rest of the >> couch_mrview/couch_index machinery. >> >> Note to self: make sure view compaction doesn’t remove deleted docs. But a >> cursory glance at couch_mrview_compactor:compact_view_btree/5 suggests no >> such thing, but we need to validate this, and if it doesn’t hold, change >> view_compation to keep deleted entries. >> >> * * * >> >> We’ll start giving this a try by forking things off in >> couch_mrview:query_all_docs/4 and pretending to call a view with a mocked >> ddoc: >> >> { >> “_id”: “_design/_access”, >> “language”: “_access” >> “views”: {} // if needed >> } // TODO see which other fields it needs >> >> We’ll try this road to see if we get to the point where we get a “view >> index not found” error, because we didn’t actually have a view index yet. >> We’ll then try and see if we can produce one. We could try the other way >> around too, building the index first and then trying to query, but the >> approach doesn’t make much of a difference. >> >> First demo working: >> https://gist.github.com/janl/20b218a3f0eafbf963ee28780261f9fc >> >> >> Next Steps: >> - make sure the startkey/endkey/descending argument handling is all >> correct and complete >> - add key un-munging, so the user/role prefix gets filtered out on reads >> - handle roles: >> - instead of querying the _access view once, we need to issue a >> multi-query, probably via #mrags.multi_get, read up on how that is used >> - then we could start thinking about by-access-seq: >> - we need access to the update-seq in >> couch_access_native_proc:map_doc, might require view protocol upgrade, or >> we have a post-process function that tags on the update-seq, we’ll see. >> - the admin/access split we’re doing in query_all_docs should probably >> happen in couch_db:changes_since/5 >> >> >> >> >> >> >> # More specification details >> >> >> Documents with in databases with _access enabled are private/admin-only by >> default, and can be made public with the special role _public >> >> TODO: shared id space or auto-prefix ids >> >> >>