Hi Vilnis,
For all performance issues, it is hard to say without all the details
because we are talking about overheads deep down the processing stack.
It makes it sensitive to the query, the data and concurrent load.
As does setup - for UI response, latency matters as well as throughput.
The only certainty is that fine-grained security does incur costs, and
that isn't a comment specific to RDF. c.f. Row level ACLs in SQL.
Inline ...
On 24/01/2022 14:42, Vilnis Termanis wrote:
Hi Andy,
Hope you're well - nice to hear from you. (responses inline)
On Sat, 22 Jan 2022 at 13:57, Andy Seaborne <a...@apache.org> wrote:
On 21/01/2022 15:26, Martynas Jusevičius wrote:
WebAccessControl ontology might be relevant here:
https://www.w3.org/wiki/WebAccessControl
We're using a request filter that controls access against
authorizations using SPARQL.
On Fri, Jan 21, 2022 at 4:13 PM Vilnis Termanis
<vilnis.terma...@iotics.com> wrote:
Hi,
Hi Vilnis,
For a SPARQL query via Fuseki, we are trying to restrict visibility of
groups of triples (each with multiple subjects) dynamically, in order
to allow for generic queries to be executed by users (instead of
providing tinned ones).
Looking at the available ACL mechanisms in Jena/Fuseki, I assume
storing each of these groups as a distinct graph might be the way
forward. (The expectation is to be able to support 10^5 or higher
number of these.)
If each graph is in the same TDB dataset, graph numbers are not much
different from any other node frequency. Millions of graphs are
possible. It's all quads. 4 Node/NodeIds.
Great, that's what I was hoping.
So it might be a way forward (details matter...)
Managing said dataset is another matter.
The description sounds a bit SOLID-like - see Martynas's comment
and -> https://inrupt.com.
I.e.: Given a user (external to Fuseki, e.g. presented via shiro via
LDAP/other), only consider triples from the set of graphs 1..N during
the query. (Where the allowed list of 1..N graphs is to be looked up
at the point of the query.)
How often is LDAP being accessed per query execution? Going off machine
is a significant cost compared to triple access. (From experience of
others, LDAP servers can be "unhelpful" - e.g. big spread in the latency
of requests based on environmental factors).
I think I could have worded that better: Given a provided (at query
time) user/Principal (which Fuseki/Jena does not have to
authenticate), only consider graphs 1..N (determined based on the
principle.)
Another factor: Location of where is the mapping of user to graphs managed.
And whether the application-logic layer is trusted to make security
decisions.
If multi-tenant, then where trust boundaries are, is going to be a factor.
details ...
(shiro is only integrated for Fuseki/webapp - it does work with
Fuseki/main but you have to add it. Current WIP should, eventually,
improve this.)
From my limited understanding, some potential routes are:
a) jena-fuseki-access - Filters triples at storage level via "TDB Quad
Filter" support in TDB.
Yes. Filtering is a hook to use. Sounds like your UC might need its own
filter code (in Java) for the policy.
However, the configuration of allowed graphs per user is static at runtime.
jena-fuseki-access is a layer on top of the filtering mechanism for the
common case of ACLs on named graphs. That layer isn't compulsory for
quad filtering. The code may be inspiration for setup of custom code.
From your perspective, if the UC is indeed only about graph-level
filtering (and not more granular), are there specific pros/cons of
implementing such a filter using the TDB quad hook VS query engine op
rewriting?
Is the former more efficient (due to being lower-level maybe?) or do
they both in the end have a very similar job - exclude matched quads
if their graphs are no in the allowed list.
Probably faster as TDB2 quad hook but "it depends".
b) jena-permissions - Extends the SPARQL query engine with an Op
...
c) Parse & re-write the query to e.g. scope it using a fixed set of
"FROM" clauses. From some minimal testing (with ~200 FROM clauses)
this does not appear to perform well (compare to a tinned query which
explicitly restricts access via knowledge of the ontologies involved).
I appreciate that maybe having a large list of FROM clauses is an
anti-pattern.
Quite likely - depends on the query complexity and numbers. There's a
hook in Fuseki query evaluation for this - did you try that or did you
do it client-side?
I've only tried it directly in the client. I presume this hook you
mention would have roughly the same perf as specifying in at the
client end?
Better, and probably still better when running as the general case of
dataset adding features over a TDB dataset.
But if the information of which 200+ is in the business logic layer
(system-trusted), as in your prototyping, it would have to be available
on the server.
In our (£job) platform, we do fine-grained security on RDF in an ABAC
style pushed as close the data as possible. Caching setup and security
decisions during execution will help.
If the query is a small amount of work, the setup overhead will be
significant but it is (roughly) a fixed overhead so a longer running
query is less impacted.
My questions are:
1) Does filtering to a set of subset of graphs (from a large set of
graphs) to restrict access sounds like a sensible thing to do? (Note
that each of these graphs would contain a set of multiple subjects -
i.e. we are not trying filter by specific predicate/object values.)
Sounds possible - "sensible" depends on the details of the intended usage.
2) Would extending either jena-fuseki-access to support the
user-graph-list lookup dynamically OR extend jena-permissions to work
at dataset level be sensible things to do?
Functionally - yes, but lots of details matter.
And user-graph-list sounds SOLID-like.
In the SOLID approach the access path is known and its the path that
decides access or not. Very different to filtering.
Our ACL definitions (for the "this user can read these graphs") are
similar to SOLID's WACL draft spec (in that, when we have more time,
we should migrate to using it). I.e. we can answer the "Can user X see
resource/graph Y" question using SPARQL - and we use this for
tinned/parameterised queries. However, if we want to allow end-users
to write their own queries but ensure the ACL still applies, then
rewriting the query (using approaches other than (c)) seems complex.
Although an implementation (e.g. Jena/Fuseki) independent generic
SPARQL solution would be great - we're considering all approaches for
performance reasons.
The data access patterns/assumptions are rather different in SOLID and I
may be behind the curve here and query over pods (or security within
pods) may have had some work done on it.
3) If the answer to either of (2) is yes - I'd be interested in
getting a better understanding of what would be involved to gauge the
size/effort of such an extension. I have had a look codebases for the
aforementioned projects, but my knowledge of TDB/ARQ/etc is very
limited. (We'd potentially be interested in taking this on, time &
priorities permitting.)
Great!
I didn't know which mailing list to send this to but I thought the
users list would probably be a better starting point.
Here is OK.
Andy
Regards,
Vilnis
