nifi-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Joey Frazee <joey.fra...@icloud.com>
Subject Re: ListSFTP incoming relationship
Date Sun, 01 Apr 2018 18:43:34 GMT
I worked on this at one point to make it "easier" (haha...) to process a very deep directory
tree with 100k+ files -- idea was to break it up into subtrees for concurrency, etc. by looping
the outgoing relationship back to input.

It ended up being being painful. Looking at the diff again the most annoying things were:

- The list processors extend AbstractListProcessor which made the change invasive. It felt
dirty to alter the already existing incoming attrs for the purpose of leaving performListing
alone so it required interface and implementation changes across other processors.

- The state problem already mentioned was a downer. The size of the state store wasn't a problem
in practice, assuming a trusted client and some vaguely small number of dirs. What was more
annoying is you have to move to directory name keys and have a state migration hook to preserve
compability. Right now there's two values stored so it's N*2 (at the time I convinced myself
those were redundant but there's a lot of edge cases that have been patched over time so dunno
for sure).

Having tried it I am on team ScanX for a second stateless processor.

If state or a last modified guard of some kind is needed, it can be implemented at the flow
level in a number of ways (DMC, LookupService, DetectDupe, etc.). This isn't possible with
ListX because it doesn't take input so embedding the last modified filter in those is way
more of a necessity.

-joey

On Apr 1, 2018, 12:06 PM -0500, Pierre Villard <pierre.villard.fr@gmail.com>, wrote:
> Hi Scott,
>
> In my opinion, based on the discussion here, I'd suggest you to implement
> the solution that you seem best to answer your needs and also taking in
> consideration all the feedback the community provided. Once you have
> something, best is to submit a pull request so that review and discussion
> can move forward on the implementation itself. I'd also recommend to file a
> JIRA with as much details as possible on what is the need, what are the
> options on the table and what is the implementation you want to propose
> (the more technical details you give, the sooner you'll get feedback for
> your code).
>
> Pierre
>
>
>
> 2018-04-01 18:40 GMT+02:00 scott <tcots8888@gmail.com>:
>
> > Okay. I guess I didn't realize how Nifi dev felt about risk tolerance. I
> > think I can work around it by adding duplicate filtering or implement some
> > other state management solution.
> > So, what's the next step?
> >
> > Scott
> >
> > On Thu, Mar 29, 2018, 10:46 AM Bryan Bende <bbende@gmail.com> wrote:
> >
> > > Scott,
> > >
> > > You are correct that the overall discussion is about allowing incoming
> > > flow files to ListSFTP.
> > >
> > > However, the previous discussion on this thread highlighted that the
> > > main reason ListSFTP currently doesn't allow incoming flow files is
> > > because of challenges when storing state.
> > >
> > > This led to the proposal of a new processor that allowed incoming flow
> > > files, but did not store state, thus avoiding the challenges mentioned
> > > above. If we were going to store state in this new processor, then
> > > we'd be back to the exact same challenges.
> > >
> > > Providing an option to turn on state also doesn't really help, because
> > > if there is an option provided to users,then the option will be used,
> > > and it needs to work when it is used.
> > >
> > > If we can come up with something that stores state and works well for
> > > all scenarios, then we aren't against it, we just need to handle the
> > > challenges highlighted by Joe's original email.
> > >
> > > Regarding some of the other ideas...
> > >
> > > The current output of ListSFTP already includes flow file attributes
> > > for each listing that include the full path, filename, last update
> > > time, owner, group, permissions, and file size.... were you thinking
> > > of something different than that?
> > >
> > > See the "Writes Attributes" section here:
> > >
> > > https://nifi.apache.org/docs/nifi-docs/components/org.
> > apache.nifi/nifi-standard-nar/1.5.0/org.apache.nifi.
> > processors.standard.ListSFTP/index.html
> > >
> > > Thanks,
> > >
> > > Bryan
> > >
> > >
> > >
> > > On Thu, Mar 29, 2018 at 12:43 PM, Andy LoPresto <alopresto@apache.org
> > > wrote:
> > > > Scott,
> > > >
> > > > I think there are two conversations going on here. You are finding the
> > > > requirements for your specific use case, and that’s great. But I echo
> > > > Bryan’s point that a community processor for this scenario should not
> > > store
> > > > state at all. Sivaprasanna’s point that given dynamic directory input,
> > > > storing state based on that can cause massive data ingestion problems
> > > still
> > > > stands.
> > > >
> > > > For your specific use case, you can prototype (or possibly even get to
> > a
> > > > stable and robust-enough point) using ExecuteScript to model the
> > behavior
> > > > you need.
> > > >
> > > > In regards to the desired output format, I would suggest a few items:
> > > >
> > > > * Avro requires a schema to be defined, and this raises the barrier to
> > > use
> > > > of the processor. Also, unless being sent to a processor that
> > understands
> > > > Avro, the result will need to be converted anyway using Record*
> > > processors.
> > > > * If the output is individual flowfiles on a 1:1 basis, each should
> > have
> > > as
> > > > many attributes populated with the parsed information as possible (i.e.
> > > > file.name, file.path, file.size, file.owner, file.permissions, etc.).
> > > This
> > > > allows for easily-consumable and routable flowfiles.
> > > > * If the output is a full directory listing, I would suggest `ls -al`
> > > type
> > > > raw text output, or JSON (arbitrary human-readable and machine-readable
> > > > format with many consuming/transforming processors).
> > > >
> > > >
> > > > Andy LoPresto
> > > > alopresto@apache.org
> > > > alopresto.apache@gmail.com
> > > > PGP Fingerprint: 70EC B3E5 98A6 5A3F D3C4 BACE 3C6E F65B 2F7D EF69
> > > >
> > > > On Mar 29, 2018, at 9:34 AM, scott <tcots8888@gmail.com> wrote:
> > > >
> > > > Sorry Bryan, but I disagree with you. Not storing state is NOT the main
> > > > point of this new processor. The main point is to allow an incoming
> > > > relationship flowfile to trigger the action, and allow variables to be
> > > used
> > > > from the attributes therein.
> > > >
> > > > I agree that if the NiFi community deems it too risky to distribute
> > this
> > > > processor with state keeping optionally available, even if the default
> > > is to
> > > > disable it, then so be it. If state is not included optionally, then
> > how
> > > > about making the output flowfile content include more than just the
> > file
> > > > names? Have it include last updated time along with the filename. If it
> > > > searches recursively, you'll want to include the path to the file also.
> > > > Maybe it would be best to output the results into a structured format,
> > > such
> > > > as AVRO? Or, maybe it would just be best to output one flowfile per
> > > remote
> > > > file found, and include updated time and fully qualified path as
> > > attributes?
> > > >
> > > > Scott
> > > >
> > > >
> > > > On 03/29/2018 04:32 AM, Bryan Bende wrote:
> > > >
> > > > The main point of the new processor is to NOT store state so that it
> > > > becomes more reasonable to allow incoming flow files.
> > > >
> > > > You could probably implement your own custom processor that does both
> > > > because you can make assumptions about how you are going to use it, but
> > > if
> > > > the NiFi community provides one then it needs to work well for all
> > > > situations, such as dynamically listing hundreds of directories, which
> > is
> > > > problematic when state is involved.
> > > >
> > > > On Thu, Mar 29, 2018 at 1:05 AM Sivaprasanna <
> > sivaprasanna246@gmail.com
> > > > wrote:
> > > >
> > > > Should we really have to have an optional state saving functionality?
> > If
> > > > the user is unaware of the implications and proceed to store the state
> > > then
> > > > what Andrew Grande mentioned will happen - possibilities of never
> > ending
> > > > stream of state information being stored. If we still go with the
> > > optional
> > > > state management approach, documentation have to be clear in explaining
> > > the
> > > > implications.
> > > >
> > > > Sivaprasanna
> > > >
> > > > On Thu, 29 Mar 2018 at 9:28 AM, scott <tcots8888@gmail.com> wrote:
> > > >
> > > > Okay. So, a new processor called "ScanSFTP", allow incoming
> > relationship
> > > > where the content of the flow file is replaced with the list of
> > matching
> > > > files from the remote directory, then the list is filtered by the usual
> > > > regex parameters like today. Optional state information is kept to
> > > > additionally filter the list of files older than the newest file
> > > > observed during the last run. Does that sound okay to everyone? If so,
> > > > what's the next step?
> > > >
> > > > Scott
> > > >
> > > >
> > > > On 03/27/2018 06:21 PM, scott wrote:
> > > >
> > > > This is a great discussion, and appreciate the interest in my problem.
> > > > I think there are workarounds if you decide not to store state, but
> > > > I'd recommend keeping it. I think state should be kept optionally,
> > > > even turned off by default. Several times I've had issues where the
> > > > state has cause me to miss files, because files get moved into the
> > > > source folder out of order, and I've wished I could turn the state
> > > > feature off.
> > > >
> > > > In my current use-case, I would not be frequently, dynamically
> > > > changing the source directory, though I can see the use-cases where it
> > > > would be. In my current use-case, I want to use an external database
> > > > table to control the configuration of all my flows. I do this by first
> > > > reading the content of the table for this particular flow ID, then
> > > > assign the result as attributes to the flowfile, essentially creating
> > > > variables I can use throughout the flow to control its behavior. This
> > > > works great with flows that initiate with HTTP or SQL, but not
> > > > ListSFTP or ListFile.
> > > >
> > > > Scott
> > > >
> > > >
> > > > On 03/27/2018 02:05 PM, Andy LoPresto wrote:
> > > >
> > > > I think Bryan’s point is a good one and when I first saw this
> > > > question (and thought of the previous times it’s been asked), my
> > > > initial response is to propose a second processor.
> > > >
> > > > Something like “ScanSFTP”/“IndexSFTP”/“SnapshotSFTP” which
operates
> > > > differently from ListSFTP — it does not maintain state, and performs
> > > > a one-time tabulation/chronicling of the state of that directory at
> > > > the given point in time.
> > > >
> > > > The responsibility to maintain and compare state across time is no
> > > > longer a requirement. There could even be a setting in the processor
> > > > to allow for “individual flowfile output” (i.e. act the same as
> > > > ListSFTP and output one flowfile per item listed) or “summary
> > > > flowfile output” where a single flowfile is generated containing the
> > > > directory listing information for all the items there. (Another
> > > > option is to output both on two different relationships).
> > > >
> > > > I think this would enable the types of workflows that users have
> > > > asked about in the past without compromising the mechanism by which
> > > > List* processors work and adding undue complexity to those processors.
> > > >
> > > > Absolutely crystal clear documentation (and a standard verb for the
> > > > new processor family) would be necessary (not only because these
> > > > processor solve different problems, but to avoid a million variants
> > > > of “I used ScanSFTP processor and it’s not tracking state”/“How
do I
> > > > provide a directory in an attribute to ListSFTP” mailing list
> > > > questions).
> > > >
> > > >
> > > > Andy LoPresto
> > > > alopresto@apache.org <mailto:alopresto@apache.org
> > > > /alopresto.apache@gmail.com <mailto:alopresto.apache@gmail.com>/
> > > > PGP Fingerprint: 70EC B3E5 98A6 5A3F D3C4 BACE 3C6E F65B 2F7D EF69
> > > >
> > > > On Mar 27, 2018, at 8:33 AM, Andrew Grande <aperepel@gmail.com
> > > > <mailto:aperepel@gmail.com>> wrote:
> > > >
> > > > The key here is that ListXXX processor maintains state. A directory
> > > > is part
> > > > of such state. Allowing arbitrary directories via an expression would
> > > > create never ending stream of new entries in the state storage,
> > > > effectively
> > > > engineering a distributed DoS attack on the NiFi node or shared ZK
> > > > quorum
> > > > (for when state is stored in there).
> > > >
> > > > Maybe if we focus on thinking about assumptions and restrictions the
> > > > processor should make to contain that risk...
> > > >
> > > > Andrew
> > > >
> > > > On Tue, Mar 27, 2018, 9:56 AM Bryan Bende <bbende@gmail.com
> > > > <mailto:bbende@gmail.com>> wrote:
> > > >
> > > > I'm not sure that would solve the problem because you'd still be
> > > > limited to one directory. What most people are asking for is the
> > > > ability to use a dynamic directory from an incoming flow file.
> > > >
> > > > I think we might be trying to fit two different use-cases into one
> > > > processor which might not make sense.
> > > >
> > > > Scenario #1... There is a directory that is constantly receiving new
> > > > data and has a significant amount of files, and I want to
> > > >
> > > > periodically
> > > >
> > > > find new files. This is what the current processors are optimized
> > > >
> > > > for.
> > > >
> > > > Scenario #2... There is a directory that is mostly static with a
> > > > moderate/small number of files, and at points in my flow I want to
> > > > dynamically perform a listing of this directory and retrieve the
> > > > files. This is more geared towards the mentality of running a
> > > > job/workflow.
> > > >
> > > >
> > > >
> > > >
> > > > On Tue, Mar 27, 2018 at 9:36 AM, Otto Fowler
> > > > <ottobackwards@gmail.com <mailto:ottobackwards@gmail.com
> > > > wrote:
> > > >
> > > > What if the changes where ‘on top of’ some base set of properties,
> > > > like
> > > > directory?
> > > > Like a filter, where if present from the incoming file will have
> > > >
> > > > the
> > > >
> > > > LIST*
> > > >
> > > > list only things
> > > > that match a name or attribute?
> > > >
> > > >
> > > >
> > > > On March 27, 2018 at 00:08:41, Joe Witt (joe.witt@gmail.com
> > > > <mailto:joe.witt@gmail.com>) wrote:
> > > >
> > > > Scott
> > > >
> > > > This idea has come up a couple of times and there is definitely
> > > > something intriguing to it. Where I think this idea stalls out
> > > >
> > > > though
> > > >
> > > > is in implementation.
> > > >
> > > > While I agree that the other List* processors might similarly
> > > >
> > > > benefit
> > > >
> > > > lets focus on ListFile. Today you tell ListFile what directory to
> > > > start looking for files in. It goes off scanning that directory for
> > > > hits and stores state about what it has already searched/seen. And
> > > >
> > > > it
> > > >
> > > > is important to keep track of how much it has already scanned
> > > >
> > > > because
> > > >
> > > > at times the search directory can be massive (100,000s of thousands
> > > >
> > > > or
> > > >
> > > > more files and directories to scan for example).
> > > >
> > > > In the proposed model the directory to be scanned could be provided
> > > > dynamically by looking at an attribute of an incoming flowfile (or
> > > > other criteria can be provided - not just the directory to scan).
> > > >
> > > > In
> > > >
> > > > this case the ListFile processor goes on scanning against that now.
> > > > What about the previous directory (or directories) it was told to
> > > > scan? Does it still track those too? What if it starts scanning the
> > > > newly provided directory, hasn't finished pulling all the data or
> > > >
> > > > new
> > > >
> > > > data is continually arriving, and it is told to switch to another
> > > > directory.
> > > >
> > > > I think if those questions can get solid answers and someone
> > > >
> > > > invests
> > > >
> > > > time in creating a PR then this could be pretty powerful. Would be
> > > > good to see a written description of the use case(s) for this too.
> > > >
> > > > Thanks
> > > > Joe
> > > >
> > > > On Mon, Mar 26, 2018 at 11:58 PM, scott <tcots8888@gmail.com
> > > > <mailto:tcots8888@gmail.com>> wrote:
> > > >
> > > > Hello Devs,
> > > >
> > > > I would like to request a feature to a major processor, ListSFTP.
> > > >
> > > > But
> > > >
> > > > before
> > > >
> > > > I do down the official road, I wanted to ask if anyone thought it
> > > > was a
> > > > terrible idea or impossible, etc. The request is to add support
> > > > for an
> > > > incoming relationship to the ListSFTP processor specifically, but
> > > >
> > > > I
> > > >
> > > > could
> > > >
> > > > see it added to many of the commonly used head processes, such as
> > > >
> > > > ListFile.
> > > >
> > > > I would envision functionality more like InvokeHTTP or
> > > > ExecuteSQL, where
> > > >
> > > > an
> > > >
> > > > incoming flow file could initiate the action, and the attributes
> > > > in the
> > > > incoming flow file could be used to configure the processor
> > > >
> > > > actions.
> > > >
> > > > It's
> > > >
> > > > the configuration aspect that most appeals to me, because it
> > > > opens it up
> > > >
> > > > to
> > > >
> > > > being centrally or dynamically configured.
> > > >
> > > > Thanks,
> > > >
> > > > Scott
> > > >
> > > >
> > > >
> > > >
> > >
> >

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message