nifi-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Matt Burgess <mattyb...@apache.org>
Subject Re: [DISCUSS] Advanced search capabilities
Date Tue, 25 Feb 2020 14:35:55 GMT
What about an in-memory (perhaps with disk persistence) representation
of the flow as a graph using Apache Tinkerpop? That would alleviate
the need for a separate graph DB while still allowing you to do graph
searches, at the cost of more memory usage. Not sure what the
footprint would look like for large in-mem flow graphs but I could do
some investigating, the AtlasReportingTask already does something
similar. For this case I wouldn't think a "normal" search would
necessarily use the graph or replace Lucene, but perhaps later we
could add a power-user graph search capability using Gremlin. I've
been looking at making the flow graph available externally (likely a
REST call), but that could be leveraged internally as well.

To Simon's Lucene filters:

1) Could we use "scope:here" and "group:myGroup" together to remove
the recursive nature of the search? Basically saying if "scope" is not
provided but "group" is, the scope defaults to "all".
2) If the term used in "group" filters is a process group name (rather
than ID), do you think there's a use case for a "parent" filter in the
case where names are provided? This would allow you, in the case of
two identically-named child PGs with different parents, to narrow the
search to the PG with the given parent.

Regards,
Matt


On Mon, Feb 24, 2020 at 12:28 PM Mike Thomsen <mikerthomsen@gmail.com> wrote:
>
> FWIW, having spent a lot of time in the last year working with graph
> database ingestion, I really don't see this story ending well for replacing
> Lucene.
>
> On Mon, Feb 24, 2020 at 5:01 PM Otto Fowler <ottobackwards@gmail.com> wrote:
>
> > +1 for the “bring your own golden hammer” approach
> >
> >
> >
> >
> > On February 24, 2020 at 11:46:14, Mike Thomsen (mikerthomsen@gmail.com)
> > wrote:
> >
> > Another thing I forgot to throw out there was that you have an issue of
> > latency if you use Janus or Neo4j. Lucene will almost certainly have
> > substantially lower latency for updating and querying the provenance data
> > if you were to do a bake off between the two to power a provenance
> > repository.
> >
> > That said, if you care more about being able to query with Cypher or
> > Gremlin than having raw performance, you could write a custom provenance
> > repository. They are pluggable.
> >
> > On Sat, Feb 22, 2020 at 7:00 AM Martin Ebert <martin.irgang@gmx.de> wrote:
> >
> > > Hi Mike,
> > > that is a fair point. You would actually raise the minimum requirements
> > of
> > > Nifi accordingly if you wanted to use a graph. As an additional
> > > application, as we are currently planning, Neo4j is nevertheless a good
> > > choice and there is nothing to be said against making it open source. The
> > > open source version of Neo4j should be sufficient for this.
> > >
> > >
> > > Mike Thomsen <mikerthomsen@gmail.com> schrieb am Sa., 22. Feb. 2020,
> > > 02:36:
> > >
> > > > Martin,
> > > >
> > > > In theory, a graph database would be superior here. Absolutely. In
> > > > practice, none of the tech out there is better than the current
> > > > Lucene-based approach in terms of ease of development and integration
> > and
> > > > low memory footprint. Adding Neo4J or JanusGraph would cause a huge
> > jump
> > > in
> > > > the minimum requirements to run NiFi. Possibly to the point where Xms
> > and
> > > > Xmx would have to start at 2GB for people getting started.
> > > >
> > > > It's been a long time since I've played with Atlas and the Atlas
> > > > integration, but if that doesn't work you can build in support for
> > Cypher
> > > > and Gremlin by adding -Pinclude-graph to a 1.10 or 1.11 build. In 1.10,
> > > one
> > > > of the NARs was overlooked in that profile, so you'd need to add it
> > back
> > > to
> > > > the profile. That was fixed in 1.11. The ExecuteGraphQuery processor
> > will
> > > > allow you to execute Cypher or Gremlin commands/scripts depending on
> > > which
> > > > controller service/driver you configure.
> > > >
> > > > On Fri, Feb 21, 2020 at 6:42 PM Martin Ebert <martin.irgang@gmx.de>
> > > wrote:
> > > >
> > > > > We still think about building a graph based search (Neo4j) in top
of
> > > > NiFi.
> > > > > Would be also fantastic to have it within NiFi.
> > > > >
> > > > > There are plenty of examples
> > > > >
> > > > >
> > > >
> > >
> >
> > https://blog.grandstack.io/using-neo4js-full-text-search-with-graphql-e3fa484de2ea
> > > > > From the idea it could go in this direction - of course much more
> > > > > rudimentary. Then one would have the possibility to have only the
> > > results
> > > > > displayed as text or to find out exploratory connections (graph
> > > layout).
> > > > > The built-in data lineage function of NiFi would also benefit from
> > the
> > > > > power of Neo4j.
> > > > >
> > > > > Simon Bence <simonbence.dev@gmail.com> schrieb am Fr., 21.
Feb.
> > 2020,
> > > > > 19:00:
> > > > >
> > > > > > Dear Community,
> > > > > >
> > > > > > In my project, I do use relatively high number of processors
and
> > > > process
> > > > > > groups. The current search function on the NiFi UI has no
> > > > capabilitites
> > > > > to
> > > > > > narrow the results based on the group, which would make the
results
> > > > more
> > > > > > relevant, so I would like to propose a possible solution. Please
if
> > > you
> > > > > > have any comment on this, do not hesitate to share it.
> > > > > >
> > > > > > The general approach would be to keep the current text box and
> > extend
> > > > the
> > > > > > server side capabilities to process search query in the similar
> > > manner
> > > > > for
> > > > > > example the Google search behaves.This extensions I would call
> > > > "filters".
> > > > > > For now I am interested in the ones I will mention below, but
I
> > > think,
> > > > it
> > > > > > is only a matter of small work for further extend the solution
with
> > > > > further
> > > > > > ones.
> > > > > >
> > > > > > In order to distinguish the filters from the rest of the search
> > > query,
> > > > I
> > > > > > propose to put them at the beginning of the query and use the
> > > > > > [a-zA-Z0-9\.]{1..n}\:[a-zA-Z0-9\.]{1..n} format. For example
a
> > filter
> > > > > might
> > > > > > look the following: lorem:ipsum
> > > > > >
> > > > > > Adding this, the search query should look like the following:
> > > > > >
> > > > > > filter1:value filter2:value rest of the query
> > > > > >
> > > > > > As for processing the filters, I suggest the following behaviour:
> > > > > >
> > > > > > - Without filters the current behaviour should be kept
> > > > > > - Everything after the filters should be handled as the search
term
> > > > > > - After the first "non filter word", anything should be considered
> > as
> > > > > part
> > > > > > of the search term (meaning: to keep the text parsing simple,
I
> > would
> > > > not
> > > > > > go in the direction to support filters at the end of the query,
> > etc.)
> > > > > > - The ordering of the filters should have no effect on the result
> > > > > > - Filter duplications should be eliminated
> > > > > > - In case a filter appears multiple times in the query, the
first
> > > > > occasion
> > > > > > will be used
> > > > > > - Unknown filters should be ignored
> > > > > > - Only adding filters will not end up with result, at least
one
> > > > character
> > > > > > must appear as search term
> > > > > >
> > > > > > Suggested filters:
> > > > > >
> > > > > > scope
> > > > > > Narrows the search based on the user's currently active process
> > > group.
> > > > > The
> > > > > > allowed values are: "all" and "here". All produces the current
> > > > behaviour,
> > > > > > thus no filtering happens, but "here" should use the current
> > process
> > > > > group
> > > > > > as "root" of the search, ignoring everything else (including
parent
> > > > > group).
> > > > > > Note: This needs a minimal frontend change, because as I did
see,
> > > > > currently
> > > > > > the current group is not sent with the search query.
> > > > > >
> > > > > > group
> > > > > > Narrows the search for a given processing group, if it exists.
The
> > > > > > behaviour is recursive, thus the result will include the contained
> > > > groups
> > > > > > as well. If it is a non-existing group, the result list should
be
> > > > empty.
> > > > > >
> > > > > > properties
> > > > > > Controls if properties values are included or not. If not provided,
> > > the
> > > > > > property values will be included. This is because in a lot of
cases
> > > > there
> > > > > > is a huge number of results come from property names.
> > > > > >
> > > > > > - Valid values for inclusion: yes, true, include, 1
> > > > > > - Valid values for exclusion: no, none, false, exclude, 0
> > > > > >
> > > > > > It is possible that the range of possible values should be limited
> > > (and
> > > > > not
> > > > > > being ambiguous), but I see a merit of "permissiveness" here
as it
> > is
> > > > > > simpler to remember.
> > > > > >
> > > > > > Also some example:
> > > > > >
> > > > > > 1.
> > > > > > scope:here properties:exclude lorem ipsum
> > > > > > This should search only in the current group (and it's children),
> > > > > excluding
> > > > > > properties and return with components containing the "lorem
ipsum"
> > > > > > expression.
> > > > > >
> > > > > > 2.
> > > > > > group:myGroup someQuery
> > > > > > This should result the finding of components with someQuery
> > > expression,
> > > > > but
> > > > > > only within the myGroup group, even if it is not the active
one.
> > > > > >
> > > > > > 3.
> > > > > > scope:all properties:include lorem
> > > > > > This should behave the same as "lorem" without filters.
> > > > > >
> > > > > > Thanks for reading, I am interested to hear your opinion!
> > > > > >
> > > > > > Kind regards,
> > > > > > Bence
> > > > > >
> > > > >
> > > >
> > >
> >

Mime
View raw message