Repository navigation
Bump parquet to 1.15.2, snappy-java and aircompressor - #2017
deepika-lalwani123 wants to merge 5 commits into
Conversation
Image scanning of gcr.io/cdapio/cdap reports jline 3.9.0 and
jetty-http 9.4.51.v20230217 inside database-plugins-*.jar. dependency:tree
shows both come from hadoop-client 3.3.6 (compile scope) ->
hadoop-yarn-client 3.3.6 (jline 3.9.0, websocket-client -> jetty-client
-> jetty-http/-util/-io 9.4.51). The main code only uses
org.apache.hadoop.{conf,io,util,mapreduce,mapred.lib.db} classes, which
come from hadoop-common / hadoop-mapreduce-client-core that CDAP provides
at runtime, so hadoop-client is now provided. This also stops packaging
the rest of the Hadoop client tree (httpclient, commons-io 2.8.0,
commons-compress 1.21, reload4j, kerby, curator, jersey-client, jackson
jaxrs 2.12.7, ...) in the database-plugins jar.
The root-managed avro-mapred 1.11.4 pulled avro-ipc and avro-ipc-jetty,
which brought velocity-engine-core 2.3 (with commons-io 2.8.0) into the
format-avro jar (and through it format-parquet / google-cloud). No plugin
uses Avro IPC, so both are excluded.
Versions:
- hadoop-client 3.3.6: compile -> provided (database-plugins/pom.xml)
- avro-ipc, avro-ipc-jetty 1.11.4: excluded from avro-mapred (root
dependencyManagement)
Bugs: Plugins side done for jline b/578444907, b/578489244, b/578492108,
b/578644778, b/578663138 and jetty-http b/578515513.
Risk: low. No main-code change; Hadoop classes are supplied by CDAP.
Downstream projects that relied on database-plugins pulling hadoop-client
transitively at compile scope must declare it themselves (none in this
repo).
Testing: JDK 8, Checkstyle on (0 violations).
mvn -DcloudBuild clean verify -pl database-plugins,format-avro,format-parquet -am:
BUILD SUCCESS; hydrator-common, format-common, format-avro,
format-parquet and database-plugins unit tests pass (database-plugins: 35
run, 1 skipped: DBConnectorTest assumes live-DB system properties). dependency:tree
-Dincludes=org.jline,org.eclipse.jetty,org.apache.avro:avro-ipc*,org.apache.velocity
shows jline/jetty only at provided/test scope and no avro-ipc or velocity.
unzip -l database-plugins jar: lib/ now only has hydrator-common, guava,
opentracing-* and common-0.1.4 (no jline, jetty, hadoop-*); format-avro
jar has no avro-ipc / velocity. DB source/sink against real MySQL /
Postgres not run (needs a database).
Image scanning of gcr.io/cdapio/cdap reports CVEs in jars packaged by hydrator-plugins: log4j-core 2.20.0 / log4j-api 2.21.1 (format-xls, core-plugins via poi), bcprov-jdk15to18 1.84 (core-plugins), commons-io 2.5 (format-delimited), json-smart 2.2.1 (transform-plugins via json-path 2.2.0) and iq80 snappy 0.2 (format-orc via orc-core 1.1.0). The root POM now imports log4j-bom and manages commons-io, json-smart and iq80 snappy, so modules no longer pin their own versions. json-path stays on 2.2.0 (2.9 would pull json-smart 2.5.x). Versions: - log4j2 (log4j-core, log4j-api) 2.20.0 / 2.21.1 -> 2.25.5 (root log4j2.version, log4j-bom import; format-xls log4j-core.version removed) - bcprov-jdk15to18 1.84 -> 1.85 (core-plugins) - commons-io 2.5 / 2.15.0 -> 2.16.1 (root commons-io.version + managed entry; format-delimited property and core-plugins / format-xls pins removed) - json-smart 2.2.1 -> 2.4.10 (root json-smart.version, managed) - iq80 snappy 0.2 -> 0.5 (root iq80-snappy.version, managed) Bugs: Plugins side done for log4j-core b/578431300, b/578491647, b/578514361, b/578514875. iq80 snappy is rows only (b/578567520 stays open through the Platform leveldb exception). Risk: low-medium. Same-major bumps. log4j 2.25.x and commons-io 2.16.1 still run on Java 8. commons-compress (1.26.0 in core-plugins / format-xls) is not touched in this PR. Testing: JDK 8, Checkstyle on (0 violations). mvn -DcloudBuild clean verify -pl core-plugins,format-xls,format-delimited,transform-plugins,format-orc -am: BUILD SUCCESS; core-plugins 328 tests (2 skipped), transform-plugins 100 (1 skipped), format-xls / format-delimited / format-orc and dependencies all pass. dependency:tree -Dincludes=org.apache.logging.log4j,org.bouncycastle,commons-io,net.minidev:json-smart,org.iq80.snappy shows only log4j 2.25.5, bcprov-jdk15to18 1.85, commons-io 2.16.1, json-smart 2.4.10 and snappy 0.5 at compile scope (smbj's bcprov-jdk15on 1.70 is runtime scope and is not embedded). unzip -l of the built jars: core-plugins lib/ has bcprov-jdk15to18-1.85, commons-io-2.16.1, log4j-api-2.25.5; format-xls has log4j-core/-api 2.25.5, commons-io-2.16.1; format-delimited commons-io-2.16.1; transform-plugins json-smart-2.4.10; format-orc snappy-0.5. Excel / ORC / SFTP pipelines in a sandbox not run.
Image scanning of gcr.io/cdapio/cdap reports org.json:json 20160212 in
transform-plugins-*.jar (3 CVEs). transform-plugins is the only module
that uses ${json.version}; only XMLToJSON uses org.json (JSONObject,
XML.toJSONObject). Kept separate from the other minor bumps because
newer org.json changes behavior: numbers are parsed as BigDecimal /
BigInteger where needed, and XML.toJSONObject keeps strings with leading
zeros and some other values that the old version coerced.
Versions:
- org.json:json 20160212 -> 20260719 (root json.version)
Bugs: none closed alone in hydrator-plugins (the 3 org.json bugs close
once WR-3b and GC-3b also land).
Risk: medium. The XMLToJSON output can differ for numeric-looking
values (for example leading zeros and very large / high-precision
numbers).
Testing: JDK 8, Checkstyle on (0 violations).
mvn -DcloudBuild clean verify -pl transform-plugins -am: BUILD SUCCESS;
transform-plugins 100 tests pass (1 skipped, ValueMapperTest, as on
the HP-2a run), including XMLToJSONConverterTest (4), XMLParserTest (8),
XMLMultiParserTest (2), JSONParserTest (8), JSONFormatterTest (2).
unzip -l transform-plugins jar: lib/json-20260719.jar (no
json-20160212). A sandbox pipeline with XMLToJSON was not run.
Image scanning of gcr.io/cdapio/cdap reports jackson-databind 2.13.4.2 and jackson-core 2.13.4 / 2.14.3 in format-avro, format-parquet (also nested in google-cloud through format-avro) and spark-plugins jars. The root POM now imports com.fasterxml.jackson:jackson-bom so every jackson module (core, databind, annotations, module-scala for the Spark tests) resolves to one version; jackson.databind.version now points at the new jackson.version property, so existing module references keep working. Versions: - jackson (core, databind, annotations, modules) 2.13.4.2 / 2.14.3 -> 2.18.12 (root jackson.version + jackson-bom import) Bugs: none closed alone; the jackson rows still have a copy shaded in parquet-jackson 1.12.0 (HP-4) and in Platform / other plugins. Risk: medium. 2.15 added stream read limits (max string length 20M, nesting depth 1000, number length 1000) and 2.18 changed some property introspection. 2.18.x still runs on Java 8. The repo has no maven-shade-plugin usage, so the shade >= 3.5 rule does not apply. Testing: JDK 8, Checkstyle on (0 violations). mvn -DcloudBuild clean verify (whole reactor: 17 modules): BUILD SUCCESS; all unit tests pass (core-plugins 328 / 2 skipped, transform-plugins 100 / 1 skipped, database-plugins 35 / 1 skipped, spark-plugins, http-plugins, solrsearch-plugins, format-* modules). dependency:tree -Dincludes=com.fasterxml.jackson.core,com.fasterxml.jackson.module,com.fasterxml.jackson.dataformat shows only 2.18.12 at compile scope. unzip -l: format-avro, format-parquet and spark-plugins lib/ contain jackson-*-2.18.12 only (format-parquet still has parquet-jackson-1.12.0, fixed in HP-4). The submodules (google-cloud, kafka, azure, ...) are separate repos and are not built here (-DcloudBuild).
Image scanning of gcr.io/cdapio/cdap reports parquet-avro 1.12.0 (1 Critical, 1 High) and the jackson-core 2.11.4 / jackson-databind 2.11.4 copy shaded inside parquet-jackson 1.12.0 in format-parquet-*.jar. parquet-avro and the other parquet modules move to 1.15.2 and parquet-jackson to 1.17.1 (it only holds the shaded jackson; 1.17.1 has a fixed copy). snappy-java (transform-plugins, root snappy.version) and aircompressor (parquet-hadoop) are aligned to the cross-track targets. Code: format-parquet ships its own copy of org.apache.parquet.avro.AvroSchemaConverter (to read INT96 as bytes). It was based on an old parquet release and shadows the 1.15.2 class, so AvroParquetWriter failed with NoSuchMethodError (AvroSchemaConverter(ParquetConfiguration)). The copy is re-based on the parquet-avro 1.15.2 source with the same INT96 change (INT96 -> Avro bytes unless parquet.avro.readInt96AsFixed is set). AvroIndexedRecordConverter / NoOpConverter are unchanged (1.12 -> 1.15 upstream changes to that class are formatting only). parquet-hadoop 1.15.2 declares parquet-jackson at runtime scope, which the bundle plugin does not embed (Embed-Dependency scope=compile), so format-parquet declares parquet-jackson explicitly to keep it in the plugin jar. Versions: - parquet-avro, -hadoop, -column, -common, -encoding, -format-structures 1.12.0 -> 1.15.2 (root parquet.version, managed) - parquet-jackson 1.12.0 -> 1.17.1 (root parquet-jackson.version, managed; explicit dependency in format-parquet) - snappy-java 1.1.2 -> 1.1.10.12 (root snappy.version) - aircompressor: new compile dependency of parquet-hadoop 1.15.2 (2.0.2) -> pinned to 2.0.3 (root aircompressor.version, managed; also replaces 0.21 from Spark's orc-core on the test classpath) Bugs: closes b/578421211 (Critical) and b/578599613 (High) alone; Plugins side done for the jackson bugs whose last hydrator copy was parquet-jackson 1.12.0. Risk: medium. New aircompressor 2.0.3, zstd-jni 1.5.6-6 and commons-pool 1.6 are now packaged (parquet-hadoop 1.15 compile deps). The managed parquet version also applies to the Spark 3.3.2 test classpath (parquet 1.12.x in Spark); all Spark-based unit tests pass. The plugin classloader loading parquet from the plugin jar on Spark 3.5.8 still needs the P2 sandbox check. All new jars are Java 8 bytecode (52). Testing: JDK 8, Checkstyle on (0 violations). mvn -DcloudBuild clean verify (whole reactor) passed for every module except format-parquet (NoSuchMethodError above, fixed in this commit); after the fix: mvn clean verify -pl format-parquet,core-plugins -am BUILD SUCCESS (format-parquet ParquetInputFormatProviderTest 3/3, core-plugins 328 / 2 skipped) and mvn clean verify -pl format-parquet -am BUILD SUCCESS with the explicit parquet-jackson. transform-plugins (100 / 1 skipped, CompressorTest / DecompressorTest with snappy), database, http, solr, spark plugins passed in the reactor run. dependency:tree -Dincludes=org.apache.parquet,io.airlift,org.xerial.snappy shows only parquet 1.15.2, parquet-jackson 1.17.1, aircompressor 2.0.3 and snappy-java 1.1.10.12 at compile scope. unzip -l format-parquet jar: lib/ has parquet-*-1.15.2, parquet-jackson-1.17.1, aircompressor-2.0.3, no parquet-*-1.12.0.
There was a problem hiding this comment.
Code Review
This pull request updates various dependency versions across multiple POM files (including Jackson, Log4j2, Snappy, and Parquet) and centralizes their management. It also updates AvroSchemaConverter.java to align with parquet-avro 1.15.2, introducing support for INT96, UUIDs, and local timestamp types. A potential NullPointerException was identified in AvroSchemaConverter.java where conf.getStrings is used without a null check before being passed to Arrays.asList.
| this.writeOldListStructure = conf.getBoolean(WRITE_OLD_LIST_STRUCTURE, WRITE_OLD_LIST_STRUCTURE_DEFAULT); | ||
| this.writeParquetUUID = conf.getBoolean(WRITE_PARQUET_UUID, WRITE_PARQUET_UUID_DEFAULT); | ||
| this.readInt96AsFixed = conf.getBoolean(READ_INT96_AS_FIXED, READ_INT96_AS_FIXED_DEFAULT); | ||
| this.pathsToInt96 = new HashSet<>(Arrays.asList(conf.getStrings(WRITE_FIXED_AS_INT96, new String[0]))); |
There was a problem hiding this comment.
If conf.getStrings returns null, passing it directly to Arrays.asList will throw a NullPointerException. Adding a null check makes this more robust.
| this.pathsToInt96 = new HashSet<>(Arrays.asList(conf.getStrings(WRITE_FIXED_AS_INT96, new String[0]))); | |
| String[] strings = conf.getStrings(WRITE_FIXED_AS_INT96, new String[0]); this.pathsToInt96 = strings == null ? Collections.emptySet() : new HashSet<>(Arrays.asList(strings)); |
There was a problem hiding this comment.
Intentional as-is — this class is a shadowed copy of upstream parquet-avro 1.15.2's AvroSchemaConverter (https://github.com/apache/parquet-java/blob/apache-parquet-1.15.2/parquet-avro/src/main/java/org/apache/parquet/avro/AvroSchemaConverter.java#L127) with only convertINT96 customized. Keeping the constructor identical to upstream to minimize diff on future Parquet upgrades (conf.getStrings also returns new String[0] whenever WRITE_FIXED_AS_INT96 is unset).
Bump parquet to 1.15.2, snappy-java and aircompressor
Repo: cdapio/hydrator-plugins
Branch: vuln/parquet-snappy-aircompressor → base develop (stacked)
Commit: bcc5ca2
Part of the CVE fixes for the CDAP 6.11.2 patch release (container image scanning of gcr.io/cdapio/cdap).
Image scanning of gcr.io/cdapio/cdap reports parquet-avro 1.12.0 (1
Critical, 1 High) and the jackson-core 2.11.4 / jackson-databind 2.11.4
copy shaded inside parquet-jackson 1.12.0 in format-parquet-*.jar.
parquet-avro and the other parquet modules move to 1.15.2 and
parquet-jackson to 1.17.1 (it only holds the shaded jackson; 1.17.1
has a fixed copy). snappy-java (transform-plugins, root
snappy.version) and aircompressor (parquet-hadoop) are aligned to the
cross-track targets.
Code: format-parquet ships its own copy of
org.apache.parquet.avro.AvroSchemaConverter (to read INT96 as bytes).
It was based on an old parquet release and shadows the 1.15.2 class,
so AvroParquetWriter failed with NoSuchMethodError
(AvroSchemaConverter(ParquetConfiguration)). The copy is re-based on
the parquet-avro 1.15.2 source with the same INT96 change (INT96 ->
Avro bytes unless parquet.avro.readInt96AsFixed is set).
AvroIndexedRecordConverter / NoOpConverter are unchanged (1.12 -> 1.15
upstream changes to that class are formatting only).
parquet-hadoop 1.15.2 declares parquet-jackson at runtime scope, which
the bundle plugin does not embed (Embed-Dependency scope=compile), so
format-parquet declares parquet-jackson explicitly to keep it in the
plugin jar.
Versions:
-format-structures 1.12.0 -> 1.15.2 (root parquet.version, managed)
managed; explicit dependency in format-parquet)
(2.0.2) -> pinned to 2.0.3 (root aircompressor.version, managed; also
replaces 0.21 from Spark's orc-core on the test classpath)
Resolves the parquet-avro 1.12.0 findings (Critical and High) on its own,
and completes the plugin fixes for the jackson findings whose last
hydrator copy was parquet-jackson 1.12.0.
Risk: medium. New aircompressor 2.0.3, zstd-jni 1.5.6-6 and commons-pool 1.6 are now
packaged (parquet-hadoop 1.15 compile deps). The managed parquet
version also applies to the Spark 3.3.2 test classpath (parquet
1.12.x in Spark); all Spark-based unit tests pass. The plugin
classloader loading parquet from the plugin jar on Spark 3.5.8 still
needs the P2 sandbox check. All new jars are Java 8 bytecode (52).
Testing: JDK 8, Checkstyle on (0 violations).
mvn -DcloudBuild clean verify (whole reactor) passed for every module
except format-parquet (NoSuchMethodError above, fixed in this commit);
after the fix: mvn clean verify -pl format-parquet,core-plugins -am
BUILD SUCCESS (format-parquet ParquetInputFormatProviderTest 3/3,
core-plugins 328 / 2 skipped) and mvn clean verify -pl format-parquet
-am BUILD SUCCESS with the explicit parquet-jackson. transform-plugins
(100 / 1 skipped, CompressorTest / DecompressorTest with snappy),
database, http, solr, spark plugins passed in the reactor run.
dependency:tree -Dincludes=org.apache.parquet,io.airlift,org.xerial.snappy
shows only parquet 1.15.2, parquet-jackson 1.17.1, aircompressor 2.0.3
and snappy-java 1.1.10.12 at compile scope. unzip -l format-parquet jar:
lib/ has parquet--1.15.2, parquet-jackson-1.17.1, aircompressor-2.0.3,
no parquet--1.12.0.