Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
Servlets Post 2.3.12, Engine 2.5.0
Description
Currently multipart POST request made to sling use the commons file upload component that parses the request fully before processing. If uploads are small they are stored in byte[], over a configurable limit they are sent to disk. This creates additional IO overhead, increases heap usage and increases upload time.
Having searched the SLing jira, and sling-dev I have failed to find an issue relating to this area, although it has been discussed in the past.
I have 2 proposals.
The SlingMain Servlet processes all requests, identifying the request type and parsing the request body. If the body is multipart the Commons File Upload library is used to process the request body in full when the SlingServletRequest is created or the first parameter is requested. To enable streaming of a request this behaviour needs to be modified. Unfortunately, processing a streamed request requires that the ultimate processor requests multipar parts in a the correct order to avoid non streaming, so a streaming behaviour will not be suitable for most POST requests and can only be used if the ultimate Servlet has been written to process a stream rather than a map of parameters.
Both proposals need to identify requests that should be processed as a stream. This identification must happen in the headers or URI as any identification later than the headers may be too late. Something like a custom header (x-uploadmode: stream) or a query string (?uploadmode=stream) or possibly a selector (/path/to/target.stream) would work and each have advantages and disadvantages.
Proposal 1
When a POST request is identified as multipart and streaming, create a LazyParameterMap that uses the Commons File Upload Streaming API (https://commons.apache.org/proper/commons-fileupload/streaming.html) to process the request on demand as parameters are requested. If parameters are requested out of sequence, do something sensible attempting to maintain streaming behaviour, but if the code really breaks streaming, throw an exception to alert servlet developer early.
Pros
- Follows a similar pattern to currently using the Servlet API.
Cons
- [] params will be hard to support when the [] is out of order, and almost impossible if the [] is an upload body.
- May not work when a request is routed incorrectly as getParameter requests will be out of streaming sequence.
Proposal 2
When a POST request is identified as multipart and streaming, create a NullParameterMap that returns null for all parameter get operations. In addition set a request Attribute containing a Iterator<Part> that allows access to the request stream in a similar way to the Commons File Upload Streaming API. Servlets that process uploads streams will use the Iterator<Part> object retrieved from the request. Part is the Servlet 3 Part https://tomcat.apache.org/tomcat-7.0-doc/servletapi/javax/servlet/http/Part.html. IIUC This API is already used in the Sling Engine and exported by a bundle.
Pros
- Won't get broken by existing getParameter calls, which all return null and do no harm to the stream.
- Far simpler implementation as the Servlet implementation has to get the request data in streaming order.
Cons
- Needs a custom Sling Upload Operation that understand how to process the Iterator<Part>
- Can't use the adaptTo mechanism on the request, as request.adaptTo(Iterator.class) doesn't make sense being too generic. Would need a new API to make this work. request.adaptTo(PartsIterator.class), which PartsIterator extends Iterator.
- Supporting the full breadth of the Sling Operation protocol in the Sling Upload Operation will require wide scale duplication of code from the ModifyOperation implementation as the ModifyOperation expects RequestProperty maps and wont work with a streamed part.
- Forces the Sling Post bundle to depend on Servlet 3 to get the Part API, requiring some patches to the existing test classes.
To support both methods a standard Servlet to handle streamed uploads would be needed, connecting the file request stream to the Resource output stream. In some cases (Oak S3 DS Async Uploads, Mongo DS) this wont entirely eliminate local disk IO, although in most cases the Resource output stream wrapps the final output stream. To maintain streaming a save operation may need to be performed for each upload to cause the request stream to be read.
If this is a duplicate issue, please link.
If you have input, please share.
Have some patches in progress, would prefer Proposal 2, as Proposal 1 looks messy at the moment.
Attachments
Attachments
Issue Links
- is broken by
-
-
- Closed
-
- links to
