I created a quick branch <https://github.com/apache/parquet-mr/compare/master...clairemcginty:parquet-mr:dict-size-repro?expand=1> to reproduce what I'm seeing -- the test shows that an Int column with cardinality 100 successfully results in a dict encoding, but an int column with cardinality 10,000 falls back and doesn't create a dict encoding. This seems like a low threshold given the 1MB dictionary page size, so I just wanted to check whether this is expected or not :)
Best, Claire On Tue, Sep 19, 2023 at 9:35 AM Claire McGinty <claire.d.mcgi...@gmail.com> wrote: > Hi, just wanted to follow up on this! > > I ran a debugger to find out why my column wasn't ending up with a > dictionary encoding and it turns out that even though > DictionaryValuesWriter#shouldFallback() > <https://github.com/apache/parquet-mr/blob/master/parquet-column/src/main/java/org/apache/parquet/column/values/dictionary/DictionaryValuesWriter.java#L117> > always returned false (dictionaryByteSize was always less than my > configured page size), DictionaryValuesWriter#isCompressionSatisfying > <https://github.com/apache/parquet-mr/blob/master/parquet-column/src/main/java/org/apache/parquet/column/values/dictionary/DictionaryValuesWriter.java#L125> > was > what was causing Parquet to switch > <https://github.com/apache/parquet-mr/blob/master/parquet-column/src/main/java/org/apache/parquet/column/values/fallback/FallbackValuesWriter.java#L75> > back to the fallback, non-dict writer. > > From what I can tell, this check compares the total byte size of > *every* element with the byte size of each *distinct* element as a kind of > proxy for encoding efficiency.... however, it seems strange that this check > can cause the writer to fall back even if the total encoded dict size is > far below the configured dictionary page size. Out of curiosity, I modified > DictionaryValuesWriter#isCompressionSatisfying > <https://github.com/apache/parquet-mr/blob/master/parquet-column/src/main/java/org/apache/parquet/column/values/dictionary/DictionaryValuesWriter.java#L125> > to > also check whether total byte size was less than dictionary max size and > re-ran my Parquet write with a local snapshot, and my file size dropped 50%. > > Best, > Claire > > On Mon, Sep 18, 2023 at 9:16 AM Claire McGinty <claire.d.mcgi...@gmail.com> > wrote: > >> Oh, interesting! I'm setting it via the >> ParquetWriter#withDictionaryPageSize method, and I do see the overall file >> size increasing when I bump the value. I'll look into it a bit more -- it >> would be helpful for some cases where the # unique values in a column is >> just over the size limit. >> >> - Claire >> >> On Fri, Sep 15, 2023 at 9:54 AM Micah Kornfield <emkornfi...@gmail.com> >> wrote: >> >>> I'll note there is also a check for encoding effectiveness [1] that could >>> come into play but I'd guess that isn't the case here. >>> >>> [1] >>> >>> https://github.com/apache/parquet-mr/blob/9b5a962df3007009a227ef421600197531f970a5/parquet-column/src/main/java/org/apache/parquet/column/values/dictionary/DictionaryValuesWriter.java#L124 >>> >>> On Fri, Sep 15, 2023 at 9:51 AM Micah Kornfield <emkornfi...@gmail.com> >>> wrote: >>> >>> > I'm glad I was looking at the right setting for dictionary size. I just >>> >> tried it out with 10x, 50x, and even total file size, though, and >>> still am >>> >> not seeing a dictionary get created. Is it possible it's bounded by >>> file >>> >> page size or some other layout option that I need to bump as well? >>> > >>> > >>> > Sorry I'm less familiar with parquet-mr, hopefully someone else to >>> chime >>> > in. If I had to guess, maybe somehow the config value isn't making it >>> to >>> > the writer (but there could also be something else at play). >>> > >>> > On Fri, Sep 15, 2023 at 9:33 AM Claire McGinty < >>> claire.d.mcgi...@gmail.com> >>> > wrote: >>> > >>> >> Thanks so much, Micah! >>> >> >>> >> I think you are using the right setting, but maybe it is possible the >>> >> > strings are still exceeding the threshold (perhaps increasing it by >>> 50x >>> >> or >>> >> > more to verify) >>> >> >>> >> >>> >> I'm glad I was looking at the right setting for dictionary size. I >>> just >>> >> tried it out with 10x, 50x, and even total file size, though, and >>> still am >>> >> not seeing a dictionary get created. Is it possible it's bounded by >>> file >>> >> page size or some other layout option that I need to bump as well? >>> >> >>> >> I haven't seen my discussion during my time in the community but >>> maybe it >>> >> > was discussed in the past. I think the main challenge here is that >>> >> pages >>> >> > are either dictionary encoded or not. I'd guess to make this >>> practical >>> >> > there would need to be a new hybrid page type, which I think it >>> might >>> >> be an >>> >> > interesting idea but quite a bit of work. Additionally, one would >>> >> likely >>> >> > need heuristics for when to potentially use the new mode versus a >>> >> complete >>> >> > fallback. >>> >> > >>> >> >>> >> Got it, thanks for the explanation! It does seem like a huge amount of >>> >> work >>> >> >>> >> >>> >> Best, >>> >> Claire >>> >> >>> >> >>> >> >>> >> On Thu, Sep 14, 2023 at 5:16 PM Micah Kornfield < >>> emkornfi...@gmail.com> >>> >> wrote: >>> >> >>> >> > > >>> >> > > - What's the heuristic for Parquet dictionary writing to succeed >>> for a >>> >> > > given column? >>> >> > >>> >> > >>> >> > >>> >> > >>> >> >>> https://github.com/apache/parquet-mr/blob/9b5a962df3007009a227ef421600197531f970a5/parquet-column/src/main/java/org/apache/parquet/column/values/dictionary/DictionaryValuesWriter.java#L117 >>> >> > >>> >> > >>> >> > > - Is that heuristic configurable at all? >>> >> > >>> >> > >>> >> > I think you are using the right setting, but maybe it is possible >>> the >>> >> > strings are still exceeding the threshold (perhaps increasing it by >>> 50x >>> >> or >>> >> > more to verify) >>> >> > >>> >> > >>> >> > > - For high-cardinality datasets, has the idea of a frequency-based >>> >> > > dictionary encoding been explored? Say, if the data follows a >>> certain >>> >> > > statistical distribution, we can create a dictionary of the most >>> >> frequent >>> >> > > values only? >>> >> > >>> >> > I haven't seen my discussion during my time in the community but >>> maybe >>> >> it >>> >> > was discussed in the past. I think the main challenge here is that >>> >> pages >>> >> > are either dictionary encoded or not. I'd guess to make this >>> practical >>> >> > there would need to be a new hybrid page type, which I think it >>> might >>> >> be an >>> >> > interesting idea but quite a bit of work. Additionally, one would >>> >> likely >>> >> > need heuristics for when to potentially use the new mode versus a >>> >> complete >>> >> > fallback. >>> >> > >>> >> > Cheers, >>> >> > Micah >>> >> > >>> >> > On Thu, Sep 14, 2023 at 12:07 PM Claire McGinty < >>> >> > claire.d.mcgi...@gmail.com> >>> >> > wrote: >>> >> > >>> >> > > Hi dev@, >>> >> > > >>> >> > > I'm running some benchmarking on Parquet read/write performance >>> and >>> >> have >>> >> > a >>> >> > > few questions about how dictionary encoding works under the hood. >>> Let >>> >> me >>> >> > > know if there's a better channel for this :) >>> >> > > >>> >> > > My test case uses parquet-avro, where I'm writing a single file >>> >> > containing >>> >> > > 5 million records. Each record has a single column, an Avro String >>> >> field >>> >> > > (Parquet binary field). I ran two configurations of base setup: >>> in the >>> >> > > first case, the string field has 5,000 possible unique values. In >>> the >>> >> > > second case, it has 50,000 unique values. >>> >> > > >>> >> > > In the first case (5k unique values), I used parquet-tools to >>> inspect >>> >> the >>> >> > > file metadata and found that a dictionary had been written: >>> >> > > >>> >> > > % parquet-tools meta testdata-case1.parquet >>> >> > > > file schema: testdata.TestRecord >>> >> > > > >>> >> > > > >>> >> > > >>> >> > >>> >> >>> -------------------------------------------------------------------------------- >>> >> > > > stringField: REQUIRED BINARY L:STRING R:0 D:0 >>> >> > > > row group 1: RC:5000001 TS:18262874 OFFSET:4 >>> >> > > > >>> >> > > > >>> >> > > >>> >> > >>> >> >>> -------------------------------------------------------------------------------- >>> >> > > > stringField: BINARY UNCOMPRESSED DO:4 FPO:38918 >>> >> > SZ:8181452/8181452/1.00 >>> >> > > > VC:5000001 ENC:BIT_PACKED,PLAIN_DICTIONARY ST:[min: 0, max: 999, >>> >> > > num_nulls: >>> >> > > > 0] >>> >> > > >>> >> > > >>> >> > > But in the second case (50k unique values), parquet-tools shows >>> that >>> >> no >>> >> > > dictionary gets created, and the file size is *much* bigger: >>> >> > > >>> >> > > % parquet-tools meta testdata-case2.parquet >>> >> > > > file schema: testdata.TestRecord >>> >> > > > >>> >> > > > >>> >> > > >>> >> > >>> >> >>> -------------------------------------------------------------------------------- >>> >> > > > stringField: REQUIRED BINARY L:STRING R:0 D:0 >>> >> > > > row group 1: RC:5000001 TS:18262874 OFFSET:4 >>> >> > > > >>> >> > > > >>> >> > > >>> >> > >>> >> >>> -------------------------------------------------------------------------------- >>> >> > > > stringField: BINARY UNCOMPRESSED DO:0 FPO:4 >>> >> SZ:43896278/43896278/1.00 >>> >> > > > VC:5000001 ENC:PLAIN,BIT_PACKED ST:[min: 0, max: 9999, >>> num_nulls: 0] >>> >> > > >>> >> > > >>> >> > > (I created a gist of my test reproduction here >>> >> > > < >>> >> >>> https://gist.github.com/clairemcginty/c3c0be85f51bc23db45a75e8d8a18806 >>> >> > >.) >>> >> > > >>> >> > > Based on this, I'm guessing there's some tip-over point after >>> which >>> >> > Parquet >>> >> > > will give up on writing a dictionary for a given column? After >>> reading >>> >> > > the Configuration >>> >> > > docs >>> >> > > < >>> >> > >>> >> >>> https://github.com/apache/parquet-mr/blob/master/parquet-hadoop/README.md >>> >> > > >, >>> >> > > I tried increasing the dictionary page size configuration 5x, >>> with the >>> >> > same >>> >> > > result (no dictionary created). >>> >> > > >>> >> > > So to summarize, my questions are: >>> >> > > >>> >> > > - What's the heuristic for Parquet dictionary writing to succeed >>> for a >>> >> > > given column? >>> >> > > - Is that heuristic configurable at all? >>> >> > > - For high-cardinality datasets, has the idea of a frequency-based >>> >> > > dictionary encoding been explored? Say, if the data follows a >>> certain >>> >> > > statistical distribution, we can create a dictionary of the most >>> >> frequent >>> >> > > values only? >>> >> > > >>> >> > > Thanks for your time! >>> >> > > - Claire >>> >> > > >>> >> > >>> >> >>> > >>> >>
