ilooner commented on a change in pull request #1296: DRILL-5365: Prevent plugin
config from changing default fs. Make DrillFileSystem Immutable.
URL: https://github.com/apache/drill/pull/1296#discussion_r203899690
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/dfs/DrillFileSystem.java
##########
@@ -65,46 +62,105 @@
import com.google.common.collect.Maps;
/**
- * DrillFileSystem is the wrapper around the actual FileSystem implementation.
+ * DrillFileSystem is the wrapper around the actual FileSystem implementation.
The {@link DrillFileSystem} is
+ * immutable.
*
* If {@link org.apache.drill.exec.ops.OperatorStats} are provided it returns
an instrumented FSDataInputStream to
* measure IO wait time and tracking file open/close operations.
*/
public class DrillFileSystem extends FileSystem implements OpenFileTracker {
static final org.slf4j.Logger logger =
org.slf4j.LoggerFactory.getLogger(DrillFileSystem.class);
private final static boolean TRACKING_ENABLED =
AssertionUtil.isAssertionsEnabled();
+ private final static DrillFileSystemCache CACHE = new DrillFileSystemCache();
+ public static final String FS_DEFAULT_NAME = "fs.default.name";
public static final String UNDERSCORE_PREFIX = "_";
public static final String DOT_PREFIX = ".";
private final ConcurrentMap<DrillFSDataInputStream, DebugStackTrace>
openedFiles = Maps.newConcurrentMap();
+ private final Configuration fsConf;
private final FileSystem underlyingFs;
private final OperatorStats operatorStats;
private final CompressionCodecFactory codecFactory;
+ private boolean initialized = false;
+
public DrillFileSystem(Configuration fsConf) throws IOException {
this(fsConf, null);
}
public DrillFileSystem(Configuration fsConf, OperatorStats operatorStats)
throws IOException {
- this.underlyingFs = FileSystem.get(fsConf);
+ // Configuration objects are mutable, and the underlying FileSystem object
may directly use a passed in Configuration.
+ // In order to avoid scenarios where a Configuration can change after a
DrillFileSystem is created, we make a copy
+ // of the Configuration.
+ this.fsConf = new Configuration(fsConf);
+ normalize(fsConf);
+
+ this.underlyingFs = CACHE.get(fsConf);
this.codecFactory = new CompressionCodecFactory(fsConf);
this.operatorStats = operatorStats;
+ this.initialized = true;
Review comment:
This is required. The case is somewhat strange and is the following:
1. The super constructor is called FileSystem()
1. The FileSystem constructor calls its super constructor Configured and
passes null for the configuration.
1. The Configured constructor tries to set a null configuration by doing
setConf(null)
The problem is we want to throw an exception if someone accidentally calls
setConf. However, we don't want setConf to throw an exception when called from
this chain of super constructor calls. Therefore we need the initialized flag
to distinguish between the two cases.
Frankly I don't like setting the flag myself, but it is required due to
design flaws in the hadoop API. The specific flaw in the api is that class
constructors should never call overridable methods.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
