我們再次檢查 Apache 產品。這次我們選擇了Solr，一個開源搜尋伺服器平台。 Solr 可讓您快速有效地搜尋資料庫和線上資源中的資訊。當面對如此複雜的任務時，即使對於經驗豐富的 Apache 開發人員來說，也很容易犯錯。在本文中，我們將討論這些類型的錯誤。
你是誰，索爾？
不久前，我們檢查了最著名的 Apache 專案之一，NetBeans IDE。我們在檢查過程中發現分析儀發出了許多有趣的警告。順便說一句，開發人員很快就注意到了他們，並在我之前提出了拉取請求:)這一次，我們決定看看他們的另一個大產品，Solr 全文搜尋平台。
Apache Solr 於 2006 年首次推出，提供了廣泛的功能，從動態叢集和資料庫整合到處理複雜格式的文件。該搜尋引擎使您能夠快速搜尋和分析網站上的信息，並且還提供在 Linux 上運行的硬體上託管搜尋伺服器的功能。 Solr 有...
長話短說，我不想用太多細節來打擾你。知道它是優化大數據的便利軟體平台就足夠了。我們為什麼不看看它的源代碼並蒐索一些有趣或不尋常的東西呢？這正是我們現在要做的事。
他們在這裡混合了一些東西
程式設計師永遠無法避免涉及混淆運算子的拼字錯誤。這是其中之一：
雷雷
此方法使用meta變數來儲存有關係統快照的資訊。在上面的程式碼片段中也檢查了meta != null。如果是這樣，則拋出 IllegalArgumentException。光是這一點就顯得很奇怪。好的，現在讓我們來看看循環體。這是呼叫 getReplicaSnapshotsForShard 的地方。鑑於這裡的meta始終為空，我們得到NullPointerException。看來開發者只是犯了一個錯字，混淆了運算子。因此，如果 meta 等於 null.
則應拋出異常
PVS-Studio 分析器充當校對員並報告偵測到的拼字錯誤：
V6008 'meta' 的空取消引用。 SolrSnapshotsTool.java 262
我想補充一點，這種混合運算符的錯誤比您想像的更常見。例如，我已經在 NetBeans 專案中看到至少兩個類似的：
雷雷
V6008 'tm' 的空取消引用。 SourcesModel.java 713
雷雷
V6008 'listeners' 的空取消引用。 MavenArtifactsImplementation.java 613
我們可能發現了新的拼字錯誤模式。我們將繼續關注。
讓我們繼續下一個片段。在本課程中，開發人員混淆了其中一個 get 方法傳回的內容。
雷雷
該類別旨在解析某個書面數學函數。有兩個屬性可以更改解析器的行為：parseMultipleSources 分析所有數值來源，parseToEnd 檢查是否應將帶有字串的函數解析到末尾。
現在讓我們來看看這些欄位的 get 和 set 方法。 parseMultipleSources 欄位在 getParseToEnd 中傳回。程式設計師在這裡混淆了應該返回什麼欄位。
分析器可以輕鬆偵測到不符合的回傳欄位：
V6091 可疑的 getter 實作。可能應該返回“parseToEnd”欄位。 FunctionQParser.java 87, FunctionQParser.java 57
以下程式碼片段中的拼字錯誤可能會導致NullPointerException.
雷雷
讓我們仔細看看：首先將value與null進行比較，然後在下一行中，對value調用length()方法。但變數可以是null！最有可能的是，開發人員應該使用 len 變數而不是呼叫 length().
由於分析器訊息，我們發現了這個拼字錯誤：
V6008 'value' 的潛在空取消引用。 IndexSizeEstimator.java 735, IndexSizeEstimator.java 736
讓我們來看另一個有拼字錯誤的程式碼片段：
雷雷
值得關注這裡的for循環。像往常一樣，程式設計師聲明循環和 idx 計數器變量，然後取得清單中的項目數。但是，有一個問題：循環的每次迭代僅採用 0 處的元素：list.get(0) 索引。
PVS-Studio 分析器偵測到此錯誤：
V6016 透過循環內的常數索引對「列表」物件的元素進行可疑存取。 AscEvaluator.java 56
The following example shows two methods with different names. However, they do the same thing.
private static List<Feature> makeFeatures(int[] featureIds) {
final List<Feature> features = new ArrayList<>();
for (final int i : featureIds) {
Map<String, Object> params = new HashMap<String, Object>();
params.put("value", i);
final Feature f = Feature.getInstance(solrResourceLoader,
ValueFeature.class.getName(), "f" + i, params);
f.setIndex(i);
features.add(f);
}
return features;
}
private static List<Feature> makeFilterFeatures(int[] featureIds) {
final List<Feature> features = new ArrayList<>();
for (final int i : featureIds) {
Map<String, Object> params = new HashMap<String, Object>();
params.put("value", i);
final Feature f = Feature.getInstance(solrResourceLoader,
ValueFeature.class.getName(), "f" + i, params);
f.setIndex(i);
features.add(f);
}
return features;
}
登入後複製
The first one creates a list of the Feature class objects. The second one, based on the name, should return a different type or filter these Features. If the FilterFeature type existed in the source code, we could assume that the developers simply made a typo. However, there's no such type. Maybe the method was copied and the developers forgot about it after copying it.
Anyway, this snippet looks very suspicious. And the analyzer proves this:
V6032 It is odd that the body of method 'makeFeatures' is fully equivalent to the body of another method 'makeFilterFeatures'. TestLTRScoringQuery.java 66, TestLTRScoringQuery.java 79
Forgot to check? Got error on track
If your gut tells you that null checks are unnecessary, don't trust it. In the code below, the "extra" check could've prevented NullPointerException.
public static Map<String, Object> postProcessCollectionJSON(
Map<String, Object> collection) {
final Map<String, Map<String, Object>> shards = collection != null // <=
? (Map<String, Map<String, Object>>)
collection.getOrDefault("shards", Collections.emptyMap())
: Collections.emptyMap();
final List<Health> healthStates = new ArrayList<>(shards.size());
shards.forEach(
....
);
collection.put("health", Health.combine(healthStates).toString()); // <=
return collection;
}
登入後複製
In the beginning of the method, the programmer checks if the collection reference is empty. If that's the case, then shards are derived from the collection. The most interesting thing is that healthStates is added to the collection at the end, regardless of whether the collection reference is empty or not.
Here's the analyzer warning for this code fragment:
V6008 Potential null dereference of 'collection'. ClusterStatus.java 303, ClusterStatus.java 335
And in the next example, the developers made an obvious mistake in the class constructor to support parallel distribution of thread work.
public class ParallelStream extends CloudSolrStream
implements Expressible {
....
private transient StreamFactory streamFactory;
public ParallelStream(String zkHost,
String collection,
String expressionString,
int workers,
StreamComparator comp
) throws IOException {
TupleStream tStream = this.streamFactory
.constructStream(expressionString); // <=
init(zkHost, collection, tStream, workers, comp);
}
....
}
登入後複製
The error lies in the first line of the constructor body. The streamFactory field is accessed here, but the field isn't initialized. The developers may have forgotten to add some logic in the constructor, or accidently may have written this line.
The PVS-Studio warning:
V6090 Field 'streamFactory' is being used before it was initialized. ParallelStream.java 61
However, they didn't forget to add a check in this method. Although, I think they put it in the wrong place.
private void createNewCollection(final String collection)
throws InterruptedException {
....
pending.add(completionService.submit(call));
while (pending != null && pending.size() > 0) {
Future<Object> future = completionService.take();
if (future == null) return;
pending.remove(future);
}
}
登入後複製
Let's look at the interaction with the pending field: first the programmer called add, then they decided to make a loop in which they gradually removed elements from the method. The most interesting thing is that they checked that pending isn't null in the loop condition. It looks very suspicious, considering that there's no variable zeroing in the loop body. Seems like they should've added a check before calling the add method as well.
The analyzer warning:
V6060 The 'pending' reference was utilized before it was verified against null. AbstractBasicDistributedZkTestBase.java 1664, AbstractBasicDistributedZkTestBase.java 1665
Lost exception
Like many modern languages, Java has exception handling feature. The most important thing is to not lose them, as it happened here.
private void doSplitShardWithRule(SolrIndexSplitter.SplitMethod splitMethod)
throws Exception {
....
try {
ZkStateReader.from(cloudClient)
.waitForState(collectionName, 30,
TimeUnit.SECONDS,
SolrCloudTestCase.activeClusterShape(1, 2));
} catch (TimeoutException e) {
new RuntimeException("Timeout waiting for
1shards and 2 replicas.", e); // <=
}
....
}
登入後複製
The error lies in the catch block: the developers created the RuntimeException object there, even added a link to the current intercepted TimeoutException and a message. But they forgot to write the throw keyword. So, the exception is never thrown.
The Lost and Found Bureau, in the form of our analyzer, found the lost exception and notified us about it:
V6006 The object was created but it is not being used. The 'throw' keyword could be missing. ShardSplitTest.java 773
How arithmetic errors interfere with testing
Does testing make software less buggy? Well, humans are the ones who write tests, and they can't help but make mistakes. This is what happened in the following example.
Public class SpellCheckCollatorTest extends SolrTestCaseJ4 {
private static final int NUM_DOCS_WITH_TERM_EVERYOTHER = 8;
private static final int NUM_DOCS = 17;
....
@Test
public void testEstimatedHitCounts() {
....
for (int val = 5; val <= 20; val++) {
String hitsXPath = xpathPrefix + "long[@name='hits']";
if (val <= NUM_DOCS_WITH_TERM_EVERYOTHER) {
int max = NUM_DOCS;
int min = (/* min collected */ val) /
(/* max docs possibly scanned */ NUM_DOCS);
hitsXPath += "[" + min + " <= . and . <= " + max + "]";
}
....
}
}
....
}
登入後複製
A string containing the min variable, which in turn is the result of dividing val by NUM_DOCS, is written to hitsXPath here. Looking closer, you can see that the maximum and minimum values of val in this fragment are 8 and 5. The NUM_DOCS value is always 17. In all cases, min is zero in integer division. Most likely, the programmer forgot to convert division arguments to real numbers and change the type of the min variable.
We found this error using a brand-new diagnostic rule in the PVS-Studio analyzer:
V6113 The '(val) / (NUM_DOCS)' expression evaluates to 0 because the absolute value of the left operand 'val' is less than the value of the right operand 'NUM_DOCS'. SpellCheckCollatorTest.java 683
Danger of checking objects by reference
The class bellow describes the equalsTo comparison method.
private static class RandomQuery extends Query {
private final long seed;
private float density;
private final List<BytesRef> docValues;
....
private boolean equalsTo(RandomQuery other) {
return seed == other.seed &&
docValues == other.docValues &&
density == other.density;
}
}
登入後複製
Comparisons of the seed and density fields almost don't cause any questions (except, perhaps, for the density field that is a real number), because the values directly written into them are considered. However, since this field has a reference type, the docValues comparison via '==' looks very dubious. This check considers only the address and not the internal state of the object.
With such defect you can miss the case when two different lists store the same values because the lists are the same, but the references are different. It seems that when the developers named the equalsTo method, they hardly meant that it should compare references rather than the internal state of objects.
The PVS-Studio analyzer warning:
V6013 Objects 'docValues' and 'other.docValues' are compared by reference. Possibly an equality comparison was intended. TestFieldCacheSortRandom.java 341
Suspicious lack of synchronization
You won't find an error in the next fragment, but it's still potentially there. How's that possible? Take a look at this code and find out why.
public abstract class CachingDirectoryFactory extends DirectoryFactory {
....
private static final Logger log = LoggerFactory.getLogger(....);
protected Map<String, CacheValue> byPathCache = new HashMap<>();
protected IdentityHashMap<Directory, CacheValue> byDirectoryCache =
new IdentityHashMap<>();
....
private void removeFromCache(CacheValue v) {
log.debug("Removing from cache: {}", v);
byDirectoryCache.remove(v.directory);
byPathCache.remove(v.path);
}
}
登入後複製
We won't be able to understand what's wrong here until we look at all the uses of the byDirectoryCache variable. In all other methods, the interaction occurs in the synchronized block. However, in the removeFromCache method, the programmer removes the collection elements outside of the synchronized blocks.
The analyzer detected this suspicious fragment:
V6102 Inconsistent synchronization of the 'byDirectoryCache' field. Consider synchronizing the field on all usages. CachingDirectoryFactory.java 92, CachingDirectoryFactory.java 228
At this point, one could say there's an error here, and the race condition could happen. However, it turns out that all removeFromCache calls are also enclosed in synchronized blocks. So, this is mostly a false positive that could be suppressed.
Although, we still can enhance this code, because there's a potential issue here. For example, when you need to use this method again, you may simply forget to enclose it in a synchronized block. Even though other methods have additional checks that the object exists in the byDirectoryCache collection, an unsynchronized call may delete an element already after the check. As a result, unnecessary actions are performed in another thread with a non-existent element of the collection, which can lead to errors in the program logic.
To protect ourselves from this, we can simply add the synchronized keyword to the removeFromCache method. So, even though there's no real error here, the static analyzer still urges us to write cleaner code.
By the way, we just recently released an article on the pitfalls of using synchronization.
Is it possible to create a few classes named the same?
In this fragment, the programmer didn't consider that classes can be renamed or declared with the same name in different packages.
private static String getFieldFlags(IndexableField f) {
IndexOptions opts = (f == null) ? null : f.fieldType().indexOptions();
StringBuilder flags = new StringBuilder();
....
flags.append((f != null && f.getClass()
.getSimpleName()
.equals("LazyField")) // <=
? FieldFlag.LAZY.getAbbreviation(): '-');
....
return flags.toString();
}
登入後複製
This is an obviously unnecessary operation that may result in an error. The f variable has the getClass method called, which returns the object type, then gets and checks the name without specifying packages. All in all, there's no error here right now. However, it can arise for two reasons.
The first one is that there may be classes with the same name in different packages. In this case, it's unclear what kind of LazyField is required, and the program will run in a different way than intended.
The second one is related to changing the class name. If the name is changed, the code won't run as intended at all. And searching for all such strings in a huge code base is very difficult. Even if you resort to searching, it's something you can just forget about.
It'd be much safer to use the instanceof operator:
flags.append(f instanceof LazyDocument.LazyField
? FieldFlag.LAZY.getAbbreviation(): '-');
登入後複製
In this case, we wouldn't need to check for null, and the code would be much shorter. The chance of an error would also decrease, if there are classes with the same name in different packages, or if the name of the class changes.
The analyzer detected a potential error and issued a warning:
V6054 Classes should not be compared by their name. LukeRequestHandler.java 247
What about documentation, though?
As a final fragment, we'll look at the following code:
@Override
public UpdateCommand clone() {
try {
return (UpdateCommand) super.clone();
} catch (CloneNotSupportedException e) {
return null; // <=
}
}
登入後複製
Let's see what's wrong with it, because everything seems fine at first glance. The analyzer informs us that returning null in the clone method is a bad idea:
V6073 It is not recommended to return null from 'clone' method. UpdateCommand.java 97
Why is it not recommended to return null from clone? It's time to consult the Java documentation:
Returns:
a clone of this instance.
Throws:
CloneNotSupportedException - if the object's class does not support the Cloneable interface. Subclasses that override the clone method can also throw this exception to indicate that an instance cannot be cloned.
The exception here indicates that the object can't be cloned. The method should return only a copy of the current object and nothing else. But why the analyzer doesn't recommend returning null from clone? It's all about further use of the code. If you constantly deviate from the recommendations in the documentation, it's difficult to catch non-standard situations.
Let's imagine a scenario where we want to use the UpdateCommand class, but the source code is unavailable, and we can't decompile it. Or we're just lazy. We can only use the built-in library with this class and focus on the interface. Our program needs us to use the clone method, so we write the following code:
try {
UpdateCommand localCopy = field.clone();
System.out.println(localCopy.toString();
} catch (CloneNotSupportedException e) {
System.out.println("Could not clone the field");
}
登入後複製
In this code, we try to catch the CloneNotSupportedException, but we can't because the exception is a NullPointerException that causes the program to crash when calling localCopy.ToString(). This comes as a complete surprise to the developer. Deviating from official recommendations can be annoying, so it's better to always follow them
Conclusion
Let's stop here and take another look at the errors we found. Most of them are the result of carelessness, but there are some that require additional thought. For example, comparing class names without considering packages, or returning null instead of throwing an exception in the clone method.
Without special development tools like static analyzers, such bugs are difficult to find, especially in projects as large as Apache Solr. If you'd like to search for such non-obvious errors in your project, you may try our static analyzer here.
By the way, Solr isn't the only Apache product we checked:
- 21 bugs in 21st version of Apache NetBeans
- Big / Bug Data: analyzing the Apache Flink source code
- Apache Hadoop code quality: production vs test
以上是搜尋查詢 Apache Solr 中的錯誤的詳細內容。更多資訊請關注PHP中文網其他相關文章！
