SNOW-1708577 Parquet V2 support for new table format #851

sfc-gh-alhuang · 2024-10-03T17:21:49Z

This PR aims to ensure consistency with the Iceberg table scanner and registration on the server side. It includes the following changes:

Added support for Parquet V2 (delta encoding) when the schema's STORAGE_SERIALIZATION_POLICY is set to OPTIMIZED.
Enabled dictionary encoding for Iceberg mode.

src/main/java/org/apache/parquet/hadoop/BdecParquetWriter.java

src/main/java/net/snowflake/ingest/streaming/internal/BlobBuilder.java

sfc-gh-hmadan · 2024-10-03T19:18:30Z

src/main/java/net/snowflake/ingest/streaming/internal/BlobBuilder.java


 if (internalParameterProvider.getEnableChunkEncryption()) {
 Pair<byte[], Integer> paddedChunk =
 padChunk(serializedChunk.chunkData, Constants.ENCRYPTION_ALGORITHM_BLOCK_SIZE_BYTES);
 byte[] paddedChunkData = paddedChunk.getFirst();
 chunkLength = paddedChunk.getSecond();

+ if (internalParameterProvider.getComputeExtendedMetadataSize()) {
+ extendedMetadataSize =


this is risky in that (a) the behavior of paddedChunkData can change and break the assumption here of no-padding-at-the-tail, (b) no guaranteed way of avoiding a corrupt value from flowing out.

lets extract extended metadata size before padding.

verify that the last four bytes are == ParquetFileWriter.MAGIC

Is there nothing on ParquetFileWriter that exposes the extendedMetadataSize?

The chunkLength returned by padChunk is the length before padding, the offset should be correct.

Added check in Utils.getExtendedMetadataSize.

Afaik the parquet file writer only log this value (ref)

src/main/java/net/snowflake/ingest/streaming/internal/BlobBuilder.java

src/main/java/net/snowflake/ingest/utils/Constants.java

sfc-gh-hmadan · 2024-10-03T19:22:05Z

src/main/java/net/snowflake/ingest/utils/Constants.java

@@ -74,6 +74,9 @@ public class Constants {
 public static final int PARQUET_MAJOR_VERSION = 1;


Remove this hardcoded version and read the actual version from one of the parquet writers / parquet config classes ?

Use major version from ParquetFileWirter.CURRENT_VERSION. Keep minor version as I cannot find one in ParquetFileWriter

src/main/java/net/snowflake/ingest/utils/ParameterProvider.java

src/main/java/net/snowflake/ingest/utils/Utils.java

src/main/java/net/snowflake/ingest/streaming/internal/InternalParameterProvider.java

src/main/java/net/snowflake/ingest/streaming/internal/EpInfo.java

sfc-gh-hmadan · 2024-10-03T19:38:30Z

src/main/java/net/snowflake/ingest/streaming/internal/AbstractRowBuffer.java

 for (Map.Entry<String, RowBufferStats> colStat : colStats.entrySet()) {
 RowBufferStats stat = colStat.getValue();
+ enableDistinctValues = stat.isEnableDistinctValue();


this looks odd to me. We are inferring an outer-class characteristic from an inner-object's state.
Lets take in a boolean alongside setDefaultValues

infact you can read this.clientBufferParameters.enableDistinctValues and don't even need an argument here.

Moved ndv flag to parameter. Cannot use this.clientBufferParameters as this is a static method called by chained of static methods.

src/main/java/net/snowflake/ingest/streaming/internal/ParquetRowBuffer.java

src/main/java/net/snowflake/ingest/streaming/internal/RowBufferStats.java

src/main/java/net/snowflake/ingest/streaming/internal/ParquetRowBuffer.java

src/main/java/net/snowflake/ingest/streaming/internal/RowBufferStats.java

sfc-gh-hmadan · 2024-10-03T19:56:46Z

src/main/java/net/snowflake/ingest/streaming/internal/RowBufferStats.java

@@ -131,6 +161,9 @@ void addStrValue(String value) {

 void addBinaryValue(byte[] valueBytes) {
 this.setCurrentMaxLength(valueBytes.length);
+ if (enableDistinctValue) {
+ distinctValues.add(new String(Base64.getEncoder().encode(valueBytes)));


ouch this will kill performance - both cpu and memory. If I'm not wrong we are already doing heavy validation of input strings, I'm hoping we can reuse that single pass on the whole string.

Another idea is to keep a Map<String, Set<Int>> here for string columns, the map key can be a CRC32 hash and the map value will be a list of row ids that contain a value with that hash. Only in the case where the Set has size > 1 do we look at the actual strings, before serialization. This is non-trivial amount of changes all over though.

Discussed offline. Lets keep NDV disabled by default. When it's enabled, skip the string ndv computation. Added a JIRA for this.

rockwotj · 2024-10-04T16:16:51Z

src/main/java/net/snowflake/ingest/utils/Constants.java

+ switch (this) {
+ case NON_ICEBERG:
+ case COMPATIBLE:
+ return ParquetProperties.WriterVersion.PARQUET_1_0;


Does this mean that non-iceberg tables (which are snowflake managed tables AFAIK) only support parquet v1?

The server-side scanner for FDN tables supports Parquet V2. This PR is specific to the Iceberg table feature and does not alter the default behavior for streaming to FDN tables.

1. Remove logic from OpenChannelResponse contract class 2. Move writerVersion defaulting to channel construction callsite in clientInternal (from channel ctor), instead of passing writerVersion=null into the channel. 3. Pass around writerVersion via RowBuffer into Flusher, instead of via the per-chunk flushContext. 4. Remove a test-only overload of ChannelInternal 5. Remove an unnecessary parameter on ChannelInternal ctor (bdecVersion) 6. Remove SerializationPolicy.NON_ICEBERG, remove the custom SerPolicy.fromName method and use Enum.valueOf that java already has

sfc-gh-hmadan · 2024-10-14T20:24:02Z

src/main/java/net/snowflake/ingest/streaming/internal/BlobBuilder.java

@@ -140,7 +141,7 @@ static <T> Blob constructBlobAndMetadata(

 if (internalParameterProvider.setIcebergSpecificFieldsInEp()) {
 chunkMetadataBuilder
- .setMajorVersion(Constants.PARQUET_MAJOR_VERSION)
+ .setMajorVersion(ParquetFileWriter.CURRENT_VERSION)


I'd much rather depend on our own constant than a third-party library's constant, I thought I had left a comment on this but don't see it anywhere :(

ok to take in next PR too, just remove the import whenever you revert this.

sfc-gh-hmadan · 2024-10-14T20:24:42Z

src/main/java/net/snowflake/ingest/streaming/internal/ClientBufferParameters.java

@@ -124,4 +124,8 @@ public Optional<Integer> getMaxRowGroups() {
 public String getParquetMessageTypeName() {
 return isIcebergMode ? PARQUET_MESSAGE_TYPE_NAME : BDEC_PARQUET_MESSAGE_TYPE_NAME;
 }
+
+ public boolean isEnableDictionaryEncoding() {
+ return isIcebergMode;


this might be dependent on storage serialization policy too, lets verify. no need to hold up PR.

sfc-gh-hmadan · 2024-10-14T20:48:51Z

src/main/java/org/apache/parquet/hadoop/BdecParquetWriter.java

@@ -43,6 +43,10 @@ public class BdecParquetWriter implements AutoCloseable {
 // Optional cap on the max number of row groups to allow per file, if this is exceeded we'll end


nit: lets rename this file to SnowflakeParquetWriter instead of BdecParquetWriter?!

sfc-gh-hmadan · 2024-10-14T20:51:11Z

src/test/java/net/snowflake/ingest/streaming/internal/datatypes/AbstractDataTypeTest.java

 databaseName = String.format("SDK_DATATYPE_COMPATIBILITY_IT_%s", getRandomIdentifier());
 conn = TestUtils.getConnection(true);
 conn.createStatement().execute(String.format("create or replace database %s;", databaseName));
 conn.createStatement().execute(String.format("use database %s;", databaseName));
 conn.createStatement().execute(String.format("use schema %s;", schemaName));

+ switch (serializationPolicy) {


nit: if (iceberg) ?

sfc-gh-hmadan

LGTM, left some minor nits/etc that can be taken with the next PR you have in flight. Lets merge this in to unblock that PR and other PRs that need to go in.

sfc-gh-alhuang requested a review from sfc-gh-hmadan October 3, 2024 17:22