https://pinot.apache.org/ logo
Join Slack
Powered by
# general
  • c
    that makes sense
  • c
    but just to close the loop on that one - if the query hits columns that are common across such segments - then there shouldn't be an issue ?
  • c
    since the 2 blocks should have the same schema at that point
  • m
    I believe so
  • c
    ok great. That answers it. Thanks a lot Mayank !
  • c
    I"ll try it out
  • m
    Sure. Let me know if you find anything contrary to what I mentioned.
  • c
    but on a related note, somehow, we got a lot of timeout exceptions :
    Copy code
    2019-02-25 07:53:17 ERROR CombineOperator:172 - Caught TimeoutException
java.util.concurrent.TimeoutException
...
    Not sure if its related - but this actually caused a perf hit on the servers. Will update with more info as I find.
  • c
    from the code, it looks like schema compatibility should fail the query (not timeout)
  • m
    I guess expensive query?
  • m
    Those might be unrelated
  • c
    well these 2 exceptions seem strongly correlated from my brief reading
  • c
    yeah lemme debug more
  • m
    Ok, let me know what you find
  • c
    sure
  • m
    The precondition is to ensure that the size to be allocated offheap for indexing is > 0. Looking at the code, it seems that this could happen if cardinality is < 10:``` public static MutableDictionary getMutableDictionary(FieldSpec.DataType dataType, boolean isOffHeapAllocation, PinotDataBufferMemoryManager memoryManager, int avgLength, int cardinality, String allocationContext) { if (isOffHeapAllocation) { // OnHeap allocation int maxOverflowSize = cardinality / 10; switch (dataType) { case INT: ```
  • m
    cc: @User
  • a
    hmm, that field is a 
    noDictionaryColumns
    and worked well so far 🤔
  • m
    The stack trace suggests it is trying to create a dictionary 
    MutableDictionaryFactory.getMutableDictionary
  • a
    yea,
    Copy code
    "noDictionaryColumns": [
      "clog_json",
    ],
    This is my config, I'm not sure why it try to create dict.
  • s
    @User the code you are pointing to is just the overflow size. That can be 0. What maybe happening is that there are segments in which the value does not appear at all.
  • s
    @User was there an issue in this regard before? I recollect seeing it before and concluded it to be a config error
  • m
    @User is this from the realtime table?
  • a
    I don’t think so. It worked fine. The only difference I could think of in the upstream pipeline we switched to json library ( from minimal json to Gson) I’m not sure that causes the issue
  • a
    Yes it’s the real-time table
  • m
    Hmm, the only way that could cause an issue is if this field is not deserialized correctly 🤔 ?
  • m
    I think there are two issues here:
    Copy code
    1. Why is the noDict config not picked up? 
2. Even if dictionary was being created, why did it see zero size?
  • s
    even though you configure no-dictionary it is possible that all conditions are not met for the consuming segment, so we create a dictinary (or, attempt to create one). Can you check if there are logs that say ? should be printed when the segment first starts consuming
  • m
    Copy code
    // Check whether to generate raw index for the column while consuming
      // Only support generating raw index on single-value non-string columns that do not have inverted index while
      // consuming. After consumption completes and the segment is built, all single-value columns can have raw index
  • m
    non-string columns
1...666768...160Latest