Uploaded image for project: 'Geode'
  1. Geode
  2. GEODE-6032

Entire object is serialized again (and again) when Delta.hasDelta() returns false and client is using PROXY Region

    XMLWordPrintableJSON

    Details

      Description

      Currently, if a cache client application is...

      1) configured with a client PROXY Region, and...
      2) the application is also using DataSerialization to de/serialize objects to/from the servers, and...
      3) the application domain objects implement the org.apache.geode.Delta interface then Apache Geode will incorrectly send the entire object (again) when Delta.hasDelta returns false.

      It is understandable that the application domain object needs to be serialized in its entirety the first time the object is sent to the server(s) (or if the object is later, subsequently removed or has expired/been evicted, and then needs to be re-added for whatever reason).

      However, once the server(s) know about the object, then only ever a "delta" should be sent, and only when Delta.hasDelta() returns true. Otherwise, if Delta.hasDelta() returns false, then most certainly, the entire object should not be sent again, otherwise it is possible for the application to enter a "race condition" scenario where the object gets "overwritten", and as a result, the application can lose data (aka "lost updates").

      If users were to change their client Region data management policy from PROXY to CACHING_PROXY then this works as expected. Apache Geode will only send updates for an object it already knows about if there is actually a "delta", otherwise Geode does nothing (that is, does not send the entire object or any delta to the server(s) since there is technically nothing to send).

      Obviously, in the CACHING_PROXY case, there is "local" state to compare against, and therefore, Geode knows about the object already, in that it "exists". It can therefore assess the object to determine if it is the same/unchanged, and not do anything in the case where Delta.hasDelta() returns false, thus the "application" informing Geode there is nothing to send.

      Clearly, in the PROXY case, there is no "local" state, and therefore, Geode does not know whether the object (already) exists on the servers or not. So, if Delta.hasDelta() returns false, it is unsure whether the objects exists or not and so decides just to send the entire object again, a "premature optimization" to be sure, which now has sacrificed "correctness", and has amplified the possible "race conditions" on the application side.

      However, this is no different than if Delta.hasDelta() returns true and the object is not yet known by the servers. When the client sends just the delta in this case, the server will send back to the client, "I don't know anything about this object for which the delta needs to be applied", and therefore, the client must turn around and send it the entire object anyway.

      So, in the PROXY case, it would be better if the client made a determination about whether the object truly exists on the server side or not, first, before arbitrarily and falsely assuming the entire object should be sent again if the Delta.hasDelta() returns false. The client simply does not know and should "verify" before sending the object.

      Obviously, this affect will performance, but is a small price to pay (and the "correct" thing to do) compared with "lost updates" and amplifying "race conditions" on the client-side.

      There is also a situation where CACHING_PROXY client Regions can even fail, and that is when copy-on-read is set to true. Hopefully, this is obvious why.

      To make matters worse, even the Javadoc explains and implies that only "pending changes" are written if they exist...

      > "Returns true if this object has pending changes it can write out."

      Of course, this doc is less than clear and very ambiguous about what exactly happens when ((hasDelta()) return false, which is "pertinent" information. But, to be sure, it is certainly not consistent in behavior when different data management policies are in effect, and most definitely not correct!

        Attachments

          Issue Links

            Activity

              People

              • Assignee:
                agingade Anilkumar Gingade
                Reporter:
                jblum John Blum
              • Votes:
                0 Vote for this issue
                Watchers:
                5 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved:

                  Time Tracking

                  Estimated:
                  Original Estimate - Not Specified
                  Not Specified
                  Remaining:
                  Remaining Estimate - 0h
                  0h
                  Logged:
                  Time Spent - 1.5h
                  1.5h