Re: [geomesa-users] GeoMesa Kryo serialization failed for SimpleFeature

[Jun Cai <joncai2012@xxxxxxxxx>]

Thanks Emilio for confirming. I have the fix and unit test ready locally. Will create a PR soon. 

Thanks,

Jun

On Wed, Oct 14, 2020 at 3:19 PM Emilio Lahr-Vivaz <elahrvivaz@xxxxxxxx> wrote:

Thanks for finding the issue, and taking the time to investigate it so thoroughly! Your solution seems correct, would you be willing to put it up as a PR? If you can share the values in the simple feature you were serializing that caused the error, we could also add a unit test to verify the fix.

Thanks,

Emilio

On 10/14/20 6:09 PM, Jun Cai wrote:

Hi All,

I am seeing serialization failure while writing features of certain length with GeoMesaFeatureWriter. Here is part of the stack trace:

Suppressed: java.lang.IllegalArgumentException: Error indexing feature '1804882973#10760608:MULTIPOLYGON
...
    at org.locationtech.geomesa.index.geotools.GeoMesaFeatureWriter$class.writeFeature(GeoMesaFeatureWriter.scala:55) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.index.geotools.GeoMesaFeatureWriter$TableFeatureWriter.writeFeature(GeoMesaFeatureWriter.scala:141) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.index.geotools.GeoMesaFeatureWriter$GeoMesaAppendFeatureWriter$class.write(GeoMesaFeatureWriter.scala:227) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.index.geotools.GeoMesaFeatureWriter$$anon$3.write(GeoMesaFeatureWriter.scala:108) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.utils.geotools.FeatureUtils$.write(FeatureUtils.scala:141) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
...
Caused by: java.lang.ArrayIndexOutOfBoundsException: 131072
    at org.locationtech.geomesa.features.kryo.impl.KryoFeatureSerialization$class.writeFeature(KryoFeatureSerialization.scala:91) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.features.kryo.impl.KryoFeatureSerialization$class.serialize(KryoFeatureSerialization.scala:42) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.features.kryo.KryoFeatureSerializer$MutableActiveSerializer.serialize(KryoFeatureSerializer.scala:78) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.index.api.WritableFeature$FeatureLevelWritableFeature$$anonfun$values$1$$anonfun$apply$1.apply(WritableFeature.scala:154) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
    at org.locationtech.geomesa.index.api.WritableFeature$FeatureLevelWritableFeature$$anonfun$values$1$$anonfun$apply$1.apply(WritableFeature.scala:154) ~[geomesa-hbase-spark-runtime_custom_2.11-2.4.1-20200904.jar:?]
...

Root Cause

With some debugging, I found the issue was in KryoFeatureSerialization at line 89 to 91:

var i = end
while (i > offset) {
    buffer(i + shift) = buffer(i)

where it moves content in Kryo Output buffer by <shift> bytes starting from end position of current feature content which is acquired from line 78:

val end = output.position()

The output.position() here is the number of bytes that have not been flushed. In the output buffer, the zero-based index of the last byte is actually (end - 1). 

This means the code copies one more byte than needed and it becomes an issue when 

output.getBuffer.length = end + shift

which returns false for the buffer expansion check at line 82 

if (output.getBuffer.length < end + shift) {

and causes the ArrayIndexOutOfBoundsException when accessing buffer(i + shift).

An example from my use case, the feature content length is 131057 (end = 131058), shift is 14, output.getBuffer.length is 131072. When serializing this feature, 

it skipped the buffer expansion and tried to access buffer(131058 + 14), which is buffer(131072), and caused the ArrayIndexOutOfBoundsException. 

Proposed Fix

Change line 89 from

var i = end

to

var i = end - 1

Per my understanding of the code and local testing, this change shouldn't cause compatibility issue in feature serialization/deserialization. But would like to get it reviewed by more folks.

Regards,

Jun Cai