aboutsummaryrefslogtreecommitdiff
path: root/launcher
diff options
context:
space:
mode:
authorMarcelo Vanzin <vanzin@cloudera.com>2015-04-14 18:51:39 -0700
committerAndrew Or <andrew@databricks.com>2015-04-14 18:51:39 -0700
commit9717389365772d218cd7c67f9a13c3440f3c6791 (patch)
treee9e38c158b4ef1be9d580e0cb4d1e489b7f8dbbc /launcher
parent6de282e2de3cb69f9b746d03fde581429248824a (diff)
downloadspark-9717389365772d218cd7c67f9a13c3440f3c6791.tar.gz
spark-9717389365772d218cd7c67f9a13c3440f3c6791.tar.bz2
spark-9717389365772d218cd7c67f9a13c3440f3c6791.zip
[SPARK-6890] [core] Fix launcher lib work with SPARK_PREPEND_CLASSES.
The fix for SPARK-6406 broke the case where sub-processes are launched when SPARK_PREPEND_CLASSES is set, because the code now would only add the launcher's build directory to the sub-process's classpath instead of the complete assembly. This patch fixes the problem by having the launch scripts stash the assembly's location in an environment variable. This is not the prettiest solution, but it avoids having to plumb that location all the way through the Worker code that launches executors. The env variable is always set by the launch scripts, so users cannot override it. Author: Marcelo Vanzin <vanzin@cloudera.com> Closes #5504 from vanzin/SPARK-6890 and squashes the following commits: 7aec921 [Marcelo Vanzin] Fix tests. ff87a60 [Marcelo Vanzin] Merge branch 'master' into SPARK-6890 31d3ce8 [Marcelo Vanzin] [SPARK-6890] [core] Fix launcher lib work with SPARK_PREPEND_CLASSES.
Diffstat (limited to 'launcher')
-rw-r--r--launcher/src/main/java/org/apache/spark/launcher/AbstractCommandBuilder.java44
-rw-r--r--launcher/src/main/java/org/apache/spark/launcher/CommandBuilderUtils.java1
-rw-r--r--launcher/src/test/java/org/apache/spark/launcher/SparkSubmitCommandBuilderSuite.java15
3 files changed, 51 insertions, 9 deletions
diff --git a/launcher/src/main/java/org/apache/spark/launcher/AbstractCommandBuilder.java b/launcher/src/main/java/org/apache/spark/launcher/AbstractCommandBuilder.java
index d8279145d8..b8f02b9611 100644
--- a/launcher/src/main/java/org/apache/spark/launcher/AbstractCommandBuilder.java
+++ b/launcher/src/main/java/org/apache/spark/launcher/AbstractCommandBuilder.java
@@ -186,12 +186,24 @@ abstract class AbstractCommandBuilder {
addToClassPath(cp, String.format("%s/core/target/jars/*", sparkHome));
}
- final String assembly = AbstractCommandBuilder.class.getProtectionDomain().getCodeSource().
- getLocation().getPath();
+ // We can't rely on the ENV_SPARK_ASSEMBLY variable to be set. Certain situations, such as
+ // when running unit tests, or user code that embeds Spark and creates a SparkContext
+ // with a local or local-cluster master, will cause this code to be called from an
+ // environment where that env variable is not guaranteed to exist.
+ //
+ // For the testing case, we rely on the test code to set and propagate the test classpath
+ // appropriately.
+ //
+ // For the user code case, we fall back to looking for the Spark assembly under SPARK_HOME.
+ // That duplicates some of the code in the shell scripts that look for the assembly, though.
+ String assembly = getenv(ENV_SPARK_ASSEMBLY);
+ if (assembly == null && isEmpty(getenv("SPARK_TESTING"))) {
+ assembly = findAssembly();
+ }
addToClassPath(cp, assembly);
- // Datanucleus jars must be included on the classpath. Datanucleus jars do not work if only
- // included in the uber jar as plugin.xml metadata is lost. Both sbt and maven will populate
+ // Datanucleus jars must be included on the classpath. Datanucleus jars do not work if only
+ // included in the uber jar as plugin.xml metadata is lost. Both sbt and maven will populate
// "lib_managed/jars/" with the datanucleus jars when Spark is built with Hive
File libdir;
if (new File(sparkHome, "RELEASE").isFile()) {
@@ -299,6 +311,30 @@ abstract class AbstractCommandBuilder {
return firstNonEmpty(childEnv.get(key), System.getenv(key));
}
+ private String findAssembly() {
+ String sparkHome = getSparkHome();
+ File libdir;
+ if (new File(sparkHome, "RELEASE").isFile()) {
+ libdir = new File(sparkHome, "lib");
+ checkState(libdir.isDirectory(), "Library directory '%s' does not exist.",
+ libdir.getAbsolutePath());
+ } else {
+ libdir = new File(sparkHome, String.format("assembly/target/scala-%s", getScalaVersion()));
+ }
+
+ final Pattern re = Pattern.compile("spark-assembly.*hadoop.*\\.jar");
+ FileFilter filter = new FileFilter() {
+ @Override
+ public boolean accept(File file) {
+ return file.isFile() && re.matcher(file.getName()).matches();
+ }
+ };
+ File[] assemblies = libdir.listFiles(filter);
+ checkState(assemblies != null && assemblies.length > 0, "No assemblies found in '%s'.", libdir);
+ checkState(assemblies.length == 1, "Multiple assemblies found in '%s'.", libdir);
+ return assemblies[0].getAbsolutePath();
+ }
+
private String getConfDir() {
String confDir = getenv("SPARK_CONF_DIR");
return confDir != null ? confDir : join(File.separator, getSparkHome(), "conf");
diff --git a/launcher/src/main/java/org/apache/spark/launcher/CommandBuilderUtils.java b/launcher/src/main/java/org/apache/spark/launcher/CommandBuilderUtils.java
index f4ebc25bdd..8028e42ffb 100644
--- a/launcher/src/main/java/org/apache/spark/launcher/CommandBuilderUtils.java
+++ b/launcher/src/main/java/org/apache/spark/launcher/CommandBuilderUtils.java
@@ -30,6 +30,7 @@ class CommandBuilderUtils {
static final String DEFAULT_MEM = "512m";
static final String DEFAULT_PROPERTIES_FILE = "spark-defaults.conf";
static final String ENV_SPARK_HOME = "SPARK_HOME";
+ static final String ENV_SPARK_ASSEMBLY = "_SPARK_ASSEMBLY";
/** Returns whether the given string is null or empty. */
static boolean isEmpty(String s) {
diff --git a/launcher/src/test/java/org/apache/spark/launcher/SparkSubmitCommandBuilderSuite.java b/launcher/src/test/java/org/apache/spark/launcher/SparkSubmitCommandBuilderSuite.java
index 626116a9e7..97043a76cc 100644
--- a/launcher/src/test/java/org/apache/spark/launcher/SparkSubmitCommandBuilderSuite.java
+++ b/launcher/src/test/java/org/apache/spark/launcher/SparkSubmitCommandBuilderSuite.java
@@ -98,7 +98,7 @@ public class SparkSubmitCommandBuilderSuite {
parser.NAME,
"appName");
- List<String> args = new SparkSubmitCommandBuilder(sparkSubmitArgs).buildSparkSubmitArgs();
+ List<String> args = newCommandBuilder(sparkSubmitArgs).buildSparkSubmitArgs();
List<String> expected = Arrays.asList("spark-shell", "--app-arg", "bar", "--app-switch");
assertEquals(expected, args.subList(args.size() - expected.size(), args.size()));
}
@@ -110,7 +110,7 @@ public class SparkSubmitCommandBuilderSuite {
parser.MASTER + "=foo",
parser.DEPLOY_MODE + "=bar");
- List<String> cmd = new SparkSubmitCommandBuilder(sparkSubmitArgs).buildSparkSubmitArgs();
+ List<String> cmd = newCommandBuilder(sparkSubmitArgs).buildSparkSubmitArgs();
assertEquals("org.my.Class", findArgValue(cmd, parser.CLASS));
assertEquals("foo", findArgValue(cmd, parser.MASTER));
assertEquals("bar", findArgValue(cmd, parser.DEPLOY_MODE));
@@ -153,7 +153,7 @@ public class SparkSubmitCommandBuilderSuite {
String deployMode = isDriver ? "client" : "cluster";
SparkSubmitCommandBuilder launcher =
- new SparkSubmitCommandBuilder(Collections.<String>emptyList());
+ newCommandBuilder(Collections.<String>emptyList());
launcher.childEnv.put(CommandBuilderUtils.ENV_SPARK_HOME,
System.getProperty("spark.test.home"));
launcher.master = "yarn";
@@ -273,10 +273,15 @@ public class SparkSubmitCommandBuilderSuite {
return contains(needle, list.split(sep));
}
- private List<String> buildCommand(List<String> args, Map<String, String> env) throws Exception {
+ private SparkSubmitCommandBuilder newCommandBuilder(List<String> args) {
SparkSubmitCommandBuilder builder = new SparkSubmitCommandBuilder(args);
builder.childEnv.put(CommandBuilderUtils.ENV_SPARK_HOME, System.getProperty("spark.test.home"));
- return builder.buildCommand(env);
+ builder.childEnv.put(CommandBuilderUtils.ENV_SPARK_ASSEMBLY, "dummy");
+ return builder;
+ }
+
+ private List<String> buildCommand(List<String> args, Map<String, String> env) throws Exception {
+ return newCommandBuilder(args).buildCommand(env);
}
}