[SPARK-14098][SQL] Generate Java code that gets a float/double value in each column of CachedBatch when DataFrame.cache() is called by kiszk · Pull Request #11956 · apache/spark

kiszk · 2016-03-25T09:02:00Z

What changes were proposed in this pull request?

This PR generates Java code to get a float/double value of each column from CachedBatch when DataFrame.cache() is called. This is done in whole stage code generation.

When DataFrame.cache() is called, data is stored as column-oriented storage (columnar cache) in CachedBatch. This PR avoid conversion from column-oriented storage to row-oriented storage.

This PR handles only float and double that are stored in a column without compression. Another PR will handle other primitive types that may be stored in a column in a compressed format. This is for ease of review by reducing the size of PR

This PR consists of two parts.

Pass data in CachedBatch to generated code by using decompress() method. CachedBatch consists of multiple ByteBuffer arrays. A ByteBuffer is just passed to generated code.
Generate code both for row-oriented storage and column-oriented storage only if
- InMemoryColumnarTableScan exists in a plan sub-tree. A decision is performed by checking an given iterator is ColumnaIterator at runtime
- Sort or join does not exist in a plan sub-tree.

This PR generates Java code for columnar cache only if types in all columns, which are accessed in operations, are primitive

This PR improves performance of aggregate sum by 3.8x - 5.2x. This benchmark is available at here

Performance results:

Java HotSpot(TM) 64-Bit Server VM 1.8.0_66-b17 on Linux 2.6.32-504.el6.x86_64
Intel(R) Xeon(R) CPU E5-2667 v2 @ 3.30GHz
Running benchmark: Double Sum with PassThrough cache
Running case: InternalRow codegen
Running case: ColumnVector codegen
Double Sum with PassThrough cache: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------
InternalRow codegen 334 / 342 47.1 21.2 1.0X
ColumnVector codegen 88 / 99 179.1 5.6 3.8X
Running benchmark: Float Sum with PassThrough cache
Running case: InternalRow codegen
Running case: ColumnVector codegen
Float Sum with PassThrough cache: Best/Avg Time(ms) Rate(M/s) Per Row(ns) Relative
-------------------------------------------------------------------------------------------
InternalRow codegen 473 / 485 66.5 15.0 1.0X
ColumnVector codegen 91 / 95 345.6 2.9 5.2X

Motivating example:

val df = sc.parallelize(0 to 9, 1).map(i => i.toFloat).toDF()
df.cache().filter("value <= 5.1").show()

Generated code

/* 001 */publicObjectgenerate(Object[] references) {
/* 002 */returnnewGeneratedIterator(references);
/* 003 */ }
/* 004 *//* 005 *//** Codegened pipeline for:/* 006 */ * Filter (value#1 <= 5.1)
/* 007 */ +- INPUT/* 008 */ */
/* 009 */finalclassGeneratedIteratorextendsorg.apache.spark.sql.execution.BufferedRowIterator {
/* 010 */privateObject[] references;
/* 011 */privatescala.collection.Iteratorinputadapter_input;
/* 012 */privateorg.apache.spark.sql.execution.metric.LongSQLMetricfilter_numOutputRows;
/* 013 */privateorg.apache.spark.sql.execution.metric.LongSQLMetricValuefilter_metricValue;
/* 014 */privateUnsafeRowfilter_result;
/* 015 */privateorg.apache.spark.sql.catalyst.expressions.codegen.BufferHolderfilter_holder;
/* 016 */privateorg.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriterfilter_rowWriter;
/* 017 */privatescala.collection.Iteratorinputadapter_input1;
/* 018 */privateintcolumnar_batchIdx;
/* 019 */privateintcolumnar_numRows;
/* 020 */privateorg.apache.spark.sql.execution.vectorized.ColumnVectorinputadapter_col0;
/* 021 */privateUnsafeRowinputadapter_result;
/* 022 */privateorg.apache.spark.sql.catalyst.expressions.codegen.BufferHolderinputadapter_holder;
/* 023 */privateorg.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriterinputadapter_rowWriter;
/* 024 */privateorg.apache.spark.sql.execution.metric.LongSQLMetricfilter_numOutputRows1;
/* 025 */privateorg.apache.spark.sql.execution.metric.LongSQLMetricValuefilter_metricValue1;
/* 026 */privateUnsafeRowfilter_result1;
/* 027 */privateorg.apache.spark.sql.catalyst.expressions.codegen.BufferHolderfilter_holder1;
/* 028 */privateorg.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriterfilter_rowWriter1;
/* 029 */privateorg.apache.spark.sql.execution.columnar.ColumnarIteratorcolumnar_itr;
/* 030 *//* 031 */publicGeneratedIterator(Object[] references) {
/* 032 */this.references = references;
/* 033 */ }
/* 034 *//* 035 */publicvoidinit(intindex, scala.collection.Iteratorinputs[]) {
/* 036 */partitionIndex = index;
/* 037 */inputadapter_input = inputs[0];
/* 038 */this.filter_numOutputRows = (org.apache.spark.sql.execution.metric.LongSQLMetric) references[0];
/* 039 */filter_metricValue = (org.apache.spark.sql.execution.metric.LongSQLMetricValue) filter_numOutputRows.localValue();
/* 040 */filter_result = newUnsafeRow(1);
/* 041 */this.filter_holder = neworg.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(filter_result, 0);
/* 042 */this.filter_rowWriter = neworg.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(filter_holder, 1);
/* 043 */inputadapter_input1 = inputs[0];
/* 044 */columnar_batchIdx = 0;
/* 045 */columnar_numRows = 0;
/* 046 */inputadapter_col0 = null;
/* 047 */inputadapter_result = newUnsafeRow(1);
/* 048 */this.inputadapter_holder = neworg.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(inputadapter_result, 0);
/* 049 */this.inputadapter_rowWriter = neworg.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(inputadapter_holder, 1);
/* 050 */this.filter_numOutputRows1 = (org.apache.spark.sql.execution.metric.LongSQLMetric) references[1];
/* 051 */filter_metricValue1 = (org.apache.spark.sql.execution.metric.LongSQLMetricValue) filter_numOutputRows1.localValue();
/* 052 */filter_result1 = newUnsafeRow(1);
/* 053 */this.filter_holder1 = neworg.apache.spark.sql.catalyst.expressions.codegen.BufferHolder(filter_result1, 0);
/* 054 */this.filter_rowWriter1 = neworg.apache.spark.sql.catalyst.expressions.codegen.UnsafeRowWriter(filter_holder1, 1);
/* 055 */columnar_itr = null;
/* 056 */ }
/* 057 *//* 058 */privatevoidprocessBatch() throwsjava.io.IOException {
/* 059 *//*** PRODUCE: Filter (value#1 <= 5.1) *//* 060 *//* 061 *//*** PRODUCE: INPUT *//* 062 *//* 063 */while (true) {
/* 064 */if (columnar_batchIdx == 0) {
/* 065 */columnar_numRows = columnar_itr.initForColumnar();
/* 066 */if (columnar_numRows < 0) {
/* 067 */cleanup();
/* 068 */break;
/* 069 */ }
/* 070 */inputadapter_col0 = columnar_itr.getColumn(0);
/* 071 */ }
/* 072 *//* 073 */while (columnar_batchIdx < columnar_numRows) {
/* 074 */intinputadapter_rowIdx = columnar_batchIdx++;
/* 075 *//*** CONSUME: Filter (value#1 <= 5.1) *//* 076 *//* 077 *//* columnVector[inputadapter_col0, inputadapter_rowIdx, float] *//* 078 */floatinputadapter_value1 = inputadapter_col0.getFloat(inputadapter_rowIdx);
/* 079 *//* 080 *//* (input[0, float] <= 5.1) *//* 081 */booleanfilter_value4 = false;
/* 082 */filter_value4 = inputadapter_value1 <= 5.1;
/* 083 */if (!filter_value4) continue;
/* 084 *//* 085 */filter_metricValue1.add(1);
/* 086 *//* 087 *//*** CONSUME: WholeStageCodegen *//* 088 *//* 089 */filter_rowWriter1.write(0, inputadapter_value1);
/* 090 */append(filter_result1);
/* 091 */if (shouldStop()) return;
/* 092 */ }
/* 093 */columnar_batchIdx = 0;
/* 094 */ }
/* 095 */ }
/* 096 *//* 097 */privatevoidprocessRow() throwsjava.io.IOException {
/* 098 *//*** PRODUCE: Filter (value#1 <= 5.1) *//* 099 *//* 100 *//*** PRODUCE: INPUT *//* 101 *//* 102 */while (inputadapter_input.hasNext()) {
/* 103 */InternalRowinputadapter_row = (InternalRow) inputadapter_input.next();
/* 104 *//*** CONSUME: Filter (value#1 <= 5.1) *//* 105 *//* 106 *//* input[0, float] *//* 107 */floatinputadapter_value = inputadapter_row.getFloat(0);
/* 108 *//* 109 *//* (input[0, float] <= 5.1) *//* 110 */booleanfilter_value = false;
/* 111 */filter_value = inputadapter_value <= 5.1;
/* 112 */if (!filter_value) continue;
/* 112 */if (!filter_value) continue;
/* 113 *//* 114 */filter_metricValue.add(1);
/* 115 *//* 116 *//*** CONSUME: WholeStageCodegen *//* 117 *//* 118 */filter_rowWriter.write(0, inputadapter_value);
/* 119 */append(filter_result);
/* 120 */if (shouldStop()) return;
/* 121 */ }
/* 122 */ }
/* 123 *//* 124 */privatevoidcleanup() {
/* 125 */inputadapter_col0 = null;
/* 126 *//* 127 */columnar_itr = null;
/* 128 */ }
/* 129 *//* 130 */protectedvoidprocessNext() throwsjava.io.IOException {
/* 131 */if ((columnar_batchIdx != 0) ||
/* 132 */ (inputadapter_input1instanceoforg.apache.spark.sql.execution.columnar.ColumnarIterator &&
/* 133 */ (columnar_itr = (org.apache.spark.sql.execution.columnar.ColumnarIterator)inputadapter_input1).isSupportColumnarCodeGen())) {
/* 134 */processBatch();
/* 135 */ } else {
/* 136 */processRow();
/* 137 */ }
/* 138 */ }
/* 139 */ }

How was this patch tested?

Tested existing test suites
added test suites for operations to dataframe generated by df.cache().

SparkQA · 2016-03-25T09:21:57Z

Test build #54165 has finished for PR 11956 at commit f61c685.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-03-25T14:17:33Z

Test build #54168 has finished for PR 11956 at commit a43f4a2.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-03-25T18:57:31Z

Test build #54178 has finished for PR 11956 at commit c26cead.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-03-31T03:47:03Z

Test build #54583 has finished for PR 11956 at commit eb6088b.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-03-31T10:25:15Z

Test build #54604 has finished for PR 11956 at commit 35a352a.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-02T07:47:29Z

Test build #54756 has finished for PR 11956 at commit d81482d.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-10T19:58:47Z

Test build #55487 has finished for PR 11956 at commit 705fe6b.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-12T19:48:44Z

Test build #55637 has finished for PR 11956 at commit ef47b8c.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-12T20:03:48Z

Test build #55639 has finished for PR 11956 at commit 19662ee.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-12T20:54:34Z

Test build #55643 has finished for PR 11956 at commit 96dbb79.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-13T20:00:31Z

Test build #55725 has finished for PR 11956 at commit e7e34cc.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-13T22:34:43Z

Test build #55733 has finished for PR 11956 at commit 76155d8.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-14T23:27:43Z

Test build #55837 has finished for PR 11956 at commit b78a1a1.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-16T23:32:31Z

Test build #56016 has finished for PR 11956 at commit c0e6bc6.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds no public classes.

kiszk · 2016-04-17T03:18:39Z

Jenkins, retest this please

SparkQA · 2016-04-17T07:08:05Z

Test build #56036 has finished for PR 11956 at commit c0e6bc6.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-17T17:21:25Z

Test build #56058 has finished for PR 11956 at commit fe2043c.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-18T21:40:34Z

Test build #56093 has finished for PR 11956 at commit 6cfa2ff.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-19T10:52:51Z

Test build #56203 has finished for PR 11956 at commit 23b6ae1.

This patch fails Spark unit tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-19T16:38:55Z

Test build #56236 has finished for PR 11956 at commit 5075cd2.

This patch fails Scala style tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-04-19T20:57:32Z

Test build #56237 has finished for PR 11956 at commit c47a492.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- class DataFrameCacheSuite extends QueryTest with SharedSQLContext

SparkQA · 2016-04-19T23:07:41Z

Test build #56255 has finished for PR 11956 at commit 8212e20.

This patch fails from timeout after a configured wait of 250m.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- class DataFrameCacheSuite extends QueryTest with SharedSQLContext

kiszk · 2016-04-20T09:09:16Z

Jenkins, retest this please

SparkQA · 2016-07-07T19:25:58Z

Test build #61919 has finished for PR 11956 at commit 54aabef.

This patch passes all tests.
This patch merges cleanly.
This patch adds no public classes.

SparkQA · 2016-07-09T11:42:38Z

Test build #62019 has finished for PR 11956 at commit fea9f34.

This patch passes all tests.
This patch merges cleanly.
This patch adds the following public classes (experimental):
- class DataFrameCacheBenchmark

kiszk · 2016-07-27T07:10:34Z

Hi @davies could you please take a look at this since Spark 2.0.0 has been successfully released?

kiszk · 2016-07-27T20:03:01Z

@rxin, could you please review this?

kiszk · 2016-08-02T19:38:29Z

@davies, could you please take a look at this since Spark 2.0.0 has been successfully released?
cc: @rxin

kiszk · 2016-08-08T05:43:46Z

@davies, I hope that you have some bandwidth to review PRs. Could you please review this, too?

kiszk · 2016-08-15T07:50:55Z

@davies could you please review this?

davies · 2016-08-15T17:32:25Z

@kiszk I'm sorry that I do not have the bandwidth to review this, https://github.com/apache/spark/pull/13899/files sounds like an easier approach (have not looked into the details), how do you think of these two?

kiszk · 2016-08-15T18:38:23Z

@davies, thank you for your comment. I hope that you will have bandwidth soon since Spark 2.0 was released.
this PR does the same thing.

In particular, generated code for reading a column is almost the same. I listed up how these two approaches does three features.

this PR does the same thing.

Generate code to read columnar storage for cache
Generate code to create columnar storage for cache
Use the new CachedBatchByte for columnar storage for cache
my PR
Generate code to read columnar storage for cache
Use the existing static code to create columnar storage for cache
Use use the existing CachedBatch that uses Array[Byte], which is wrapped by ByteBufferColumnVector, for columnar storage for cache

I like to simplify my PR by using the idea in the PR. For example, I can throw away new files ByteBufferColumnVector.java by implementing cache using ColumnarBatch and PassThroughSuite.scala by generating code.

I have the following questions:

Do we generate code to create columnar storage for cache?
Do we use the existing CachedBatch or ColumnarBatch for cache?
In this implementation, how a cache content in ColumnarBatch will be serialized when it must be flushed into a disk?
What test cases were failed in the PR? Links to test results are not valid now.
Will we support compression scheme in the future if we use ColumnarBatch?

What do you think?

kiszk · 2016-08-18T18:51:05Z

@davies We would appreciate it if you would make comments on my questions to share how we implement a part of columnar.

kiszk · 2016-08-22T08:05:12Z

@davies Would it be possible to share your opinions regarding these design questions among us?

a-roberts · 2016-08-22T08:47:46Z

@hvanhovell @marmbrus @srowen I see this PR has been open since the 25th of March and provides substantial performance improvements as mentioned above without introducing functional regressions, as leading SQL/community members what do you guys think?

srowen · 2016-08-22T08:58:26Z

I'm not qualified to comment as I tend to ignore SQL unless it's a simple and easy to evaluate change.

kiszk · 2016-08-29T14:25:27Z

@davies Could you please share your great opinions regarding these design questions among our community while we know you are busy?

davies · 2016-08-29T17:55:28Z

@kiszk The current implementation use ByteBuffer and smart compression algorithms, it too slow to build the in-memory cache, make it useless. So we'd like to improve the performance of building phase also. This PR is built for current representation, which may be throwed away in future, so I'd not rush to merge this PR, or spent to much time to review the details.

PR #13899 could be in the right direction, but need to double-check that by more benchmarks. In order to have better memory efficiency, we could use MEMORY_AND_SER storage level and compress the underlying array with LZ4 when serializing the ColumnVector.

davies · 2016-08-29T17:55:39Z

cc @rxin

kiszk · 2016-08-29T20:21:23Z

@davies Thank you for sharing your valuable thought. I understand the future direction. I will implement this direction by using this PR or another PR. While the future roadmap may consume more memory, the performance will be improved. The compression will be applied only when ColumnVector is serialized into disk. Am I correct?

What benchmark programs do you want to use to double-check this roadmap?

davies · 2016-08-29T20:47:04Z

We could compress them in memory with MEMORY_AND_DISK_SER, this could be controlled by a flag.

kiszk · 2016-08-30T02:28:41Z

Thank you for your clarification. Here is a good summary for me.
We may add property like spark.sql.cache.compression.codec to specify a compression method.

kiszk · 2016-08-30T02:32:09Z

@davies, could you please rerun Jenkins for PR #13899? I would like to know which tests were failed by the current implementation.

kiszk · 2016-09-02T08:19:43Z

I saw PR #13899. I understood there are two design points for now. Hopefully, no more points :)

CachedBatch must be serializable.
According to my experiments, we should make a class for a CachedBatch simple.

For 1, current ColumnarBatch is not serializable.
For 2, at this line, .next() returns CachedBatch, and then += involves call SizeEstimator.estimate(). If this method takes longer time for a complicated CacheBatch class, the performance is degraded. One of example is auto_join?? in this test.

kiszk · 2016-09-21T02:56:02Z

Current PR #13899 does not support a case that an element is null.
For now, I am working for the followings:

Support all of types (seems not easy to support UnsafeArrayData in ColumnVector efficiently)
Make structures serializable
Support null cases

SparkQA · 2017-01-10T22:35:52Z

Test build #71153 has finished for PR 11956 at commit fea9f34.

This patch passes all tests.
This patch does not merge cleanly.
This patch adds the following public classes (experimental):
class DataFrameCacheBenchmark

SparkQA · 2017-02-28T23:52:37Z

Test build #73611 has finished for PR 11956 at commit fea9f34.

This patch fails PySpark unit tests.
This patch does not merge cleanly.
This patch adds the following public classes (experimental):
class DataFrameCacheBenchmark

gatorsmile · 2017-10-28T04:37:33Z

@kiszk Is this still the issue? I knew you are working on the related part now.

kiszk · 2017-10-28T08:06:31Z

Thank you for pointing it out. #18747 implemented this feature.

kiszk force-pushed the SPARK-14098 branch from c26cead to eb6088bCompare March 31, 2016 03:10

kiszk force-pushed the SPARK-14098 branch from d81482d to 705fe6bCompare April 10, 2016 19:22

kiszk force-pushed the SPARK-14098 branch from 705fe6b to ef47b8cCompare April 12, 2016 19:46

kiszk changed the title ~~[SPARK-14098][SQL][WIP] Generate Java code that gets a value in each column of CachedBatch when DataFrame.cache() is called~~[SPARK-14098][SQL] Generate Java code that gets a value in each column of CachedBatch when DataFrame.cache() is calledApr 17, 2016

kiszk force-pushed the SPARK-14098 branch from 4598cb5 to 5454065Compare April 18, 2016 05:59

kiszk force-pushed the SPARK-14098 branch from 6cfa2ff to 23b6ae1Compare April 19, 2016 07:18

kiszk force-pushed the SPARK-14098 branch from c47a492 to 8212e20Compare April 19, 2016 18:53

kiszk added 2 commits July 9, 2016 18:54

remove unnecessary final attribute
4775db2

update
fea9f34

kiszk mentioned this pull request Sep 23, 2016
[WIP][SPARK-14098][SQL] Generate Java code to build CachedColumnarBatch and get values from CachedColumnarBatch when DataFrame.cache() is called #15219
Closed

kiszk closed this Oct 28, 2017

Uh oh!

Conversation

kiszk commented Mar 25, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changes were proposed in this pull request?

How was this patch tested?

Uh oh!

SparkQA commented Mar 25, 2016

Uh oh!

SparkQA commented Mar 25, 2016

Uh oh!

SparkQA commented Mar 25, 2016

Uh oh!

SparkQA commented Mar 31, 2016

Uh oh!

SparkQA commented Mar 31, 2016

Uh oh!

SparkQA commented Apr 2, 2016

Uh oh!

SparkQA commented Apr 10, 2016

Uh oh!

SparkQA commented Apr 12, 2016

Uh oh!

SparkQA commented Apr 12, 2016

Uh oh!

SparkQA commented Apr 12, 2016

Uh oh!

SparkQA commented Apr 13, 2016

Uh oh!

SparkQA commented Apr 13, 2016

Uh oh!

SparkQA commented Apr 14, 2016

Uh oh!

SparkQA commented Apr 16, 2016

Uh oh!

kiszk commented Apr 17, 2016

Uh oh!

SparkQA commented Apr 17, 2016

Uh oh!

SparkQA commented Apr 17, 2016

Uh oh!

SparkQA commented Apr 18, 2016

Uh oh!

SparkQA commented Apr 19, 2016

Uh oh!

SparkQA commented Apr 19, 2016

Uh oh!

SparkQA commented Apr 19, 2016

Uh oh!

SparkQA commented Apr 19, 2016

Uh oh!

kiszk commented Apr 20, 2016

Uh oh!

SparkQA commented Jul 7, 2016

Uh oh!

SparkQA commented Jul 9, 2016

Uh oh!

kiszk commented Jul 27, 2016

Uh oh!

kiszk commented Jul 27, 2016

Uh oh!

kiszk commented Aug 2, 2016

Uh oh!

kiszk commented Aug 8, 2016

Uh oh!

kiszk commented Aug 15, 2016

Uh oh!

davies commented Aug 15, 2016

Uh oh!

kiszk commented Aug 15, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kiszk commented Aug 18, 2016• edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

kiszk commented Aug 22, 2016

Uh oh!

a-roberts commented Aug 22, 2016

Uh oh!

srowen commented Aug 22, 2016

kiszk commented Mar 25, 2016•
edited
Loading

kiszk commented Aug 15, 2016•
edited
Loading

kiszk commented Aug 18, 2016•
edited
Loading

kiszk commented Aug 29, 2016•
edited
Loading

kiszk commented Sep 21, 2016•
edited
Loading