(converter) Fix rare sentence extractor bug

It was caused by non-thread safe concurrent memory access in SentenceExtractor.

code/libraries/language-processing/src/main/java/nu/marginalia/language/sentence/SentenceExtractor.java

@@ -22,6 +22,11 @@
 import java.io.InputStream;
 import java.util.*;
 
+/**  This class still isn't thread safe!!  If you use it concurrently, it won't throw exceptions,
+ * it will just haunt your code and cause unpredictable mysterious errors.
+ *
+ * Use {@link ThreadLocalSentenceExtractorProvider} instead to avoid falling into the twilight zone!
+ */
 public class SentenceExtractor {
 
     private SentenceDetectorME sentenceDetector;

code/libraries/language-processing/src/main/java/nu/marginalia/language/sentence/ThreadLocalSentenceExtractorProvider.java

@@ -0,0 +1,19 @@
+package nu.marginalia.language.sentence;
+
+import com.google.inject.Inject;
+import com.google.inject.Singleton;
+import nu.marginalia.LanguageModels;
+
+@Singleton
+public class ThreadLocalSentenceExtractorProvider {
+    private final ThreadLocal<SentenceExtractor> sentenceExtractorThreadLocal;
+
+    @Inject
+    public ThreadLocalSentenceExtractorProvider(LanguageModels languageModels) {
+        sentenceExtractorThreadLocal = ThreadLocal.withInitial(() -> new SentenceExtractor(languageModels));
+    }
+
+    public SentenceExtractor get() {
+        return sentenceExtractorThreadLocal.get();
+    }
+}

code/processes/converting-process/src/main/java/nu/marginalia/converting/processor/plugin/HtmlDocumentProcessorPlugin.java

@@ -10,6 +10,7 @@
 import nu.marginalia.converting.processor.logic.links.LinkProcessor;
 import nu.marginalia.converting.processor.plugin.specialization.*;
 import nu.marginalia.language.model.DocumentLanguageData;
+import nu.marginalia.language.sentence.ThreadLocalSentenceExtractorProvider;
 import nu.marginalia.model.crawl.HtmlFeature;
 import nu.marginalia.link_parser.LinkParser;
 import nu.marginalia.crawling.model.CrawledDocument;
@@ -44,7 +45,6 @@ public class HtmlDocumentProcessorPlugin extends AbstractDocumentProcessorPlugin
     private final Logger logger = LoggerFactory.getLogger(getClass());
     private final double minDocumentQuality;
 
-    private final SentenceExtractor sentenceExtractor;
     private final FeatureExtractor featureExtractor;
     private final TitleExtractor titleExtractor;
     private final DocumentKeywordExtractor keywordExtractor;
@@ -59,6 +59,7 @@ public class HtmlDocumentProcessorPlugin extends AbstractDocumentProcessorPlugin
     private static final LinkParser linkParser = new LinkParser();
     private static final FeedExtractor feedExtractor = new FeedExtractor(linkParser);
 
+    private final ThreadLocalSentenceExtractorProvider sentenceExtractorProvider;
     private final HtmlProcessorSpecializations htmlProcessorSpecializations;
 
     @Inject
@@ -73,13 +74,13 @@ public HtmlDocumentProcessorPlugin(
             DocumentLengthLogic documentLengthLogic,
             MetaRobotsTag metaRobotsTag,
             DocumentGeneratorExtractor documentGeneratorExtractor,
+            ThreadLocalSentenceExtractorProvider sentenceExtractorProvider,
             HtmlProcessorSpecializations specializations)
     {
         super(languageFilter);
 
         this.documentLengthLogic = documentLengthLogic;
         this.minDocumentQuality = minDocumentQuality;
-        this.sentenceExtractor = sentenceExtractor;
         this.featureExtractor = featureExtractor;
 
         this.titleExtractor = titleExtractor;
@@ -88,6 +89,7 @@ public HtmlDocumentProcessorPlugin(
         this.metaRobotsTag = metaRobotsTag;
 
         this.documentGeneratorExtractor = documentGeneratorExtractor;
+        this.sentenceExtractorProvider = sentenceExtractorProvider;
         this.htmlProcessorSpecializations = specializations;
     }
 
@@ -122,7 +124,8 @@ public DetailsWithWords createDetails(CrawledDocument crawledDocument)
             throw new DisqualifiedException(DisqualificationReason.IRRELEVANT);
         }
 
-        DocumentLanguageData dld = sentenceExtractor.extractSentences(specialization.prune(doc));
+        DocumentLanguageData dld =
+                sentenceExtractorProvider.get().extractSentences(specialization.prune(doc));
 
         checkDocumentLanguage(dld);
 

code/processes/converting-process/src/main/java/nu/marginalia/converting/sideload/SideloadSourceFactory.java

@@ -6,7 +6,7 @@
 import nu.marginalia.converting.sideload.encyclopedia.EncyclopediaMarginaliaNuSideloader;
 import nu.marginalia.converting.sideload.stackexchange.StackexchangeSideloader;
 import nu.marginalia.keyword.DocumentKeywordExtractor;
-import nu.marginalia.language.sentence.SentenceExtractor;
+import nu.marginalia.language.sentence.ThreadLocalSentenceExtractorProvider;
 
 import java.io.IOException;
 import java.nio.file.Files;
@@ -17,19 +17,19 @@
 public class SideloadSourceFactory {
     private final Gson gson;
     private final SideloaderProcessing sideloaderProcessing;
-    private final SentenceExtractor sentenceExtractor;
+    private final ThreadLocalSentenceExtractorProvider sentenceExtractorProvider;
     private final DocumentKeywordExtractor documentKeywordExtractor;
     private final DirtreeSideloaderFactory dirtreeSideloaderFactory;
 
     @Inject
     public SideloadSourceFactory(Gson gson,
                                  SideloaderProcessing sideloaderProcessing,
-                                 SentenceExtractor sentenceExtractor,
+                                 ThreadLocalSentenceExtractorProvider sentenceExtractorProvider,
                                  DocumentKeywordExtractor documentKeywordExtractor,
                                  DirtreeSideloaderFactory dirtreeSideloaderFactory) {
         this.gson = gson;
         this.sideloaderProcessing = sideloaderProcessing;
-        this.sentenceExtractor = sentenceExtractor;
+        this.sentenceExtractorProvider = sentenceExtractorProvider;
         this.documentKeywordExtractor = documentKeywordExtractor;
         this.dirtreeSideloaderFactory = dirtreeSideloaderFactory;
     }
@@ -48,7 +48,7 @@ public Collection<? extends SideloadSource> sideloadStackexchange(Path pathToDbF
             return dirs
                 .filter(Files::isRegularFile)
                 .filter(f -> f.toFile().getName().endsWith(".db"))
-                .map(dbFile -> new StackexchangeSideloader(dbFile, sentenceExtractor, documentKeywordExtractor))
+                .map(dbFile -> new StackexchangeSideloader(dbFile, sentenceExtractorProvider, documentKeywordExtractor))
                 .toList();
         }
     }

code/processes/converting-process/src/main/java/nu/marginalia/converting/sideload/SideloaderProcessing.java

@@ -8,7 +8,6 @@
 import nu.marginalia.model.EdgeUrl;
 import nu.marginalia.model.crawl.UrlIndexingState;
 import nu.marginalia.model.idx.WordFlags;
-import nu.marginalia.model.idx.WordMetadata;
 
 import java.net.URISyntaxException;
 import java.time.LocalDateTime;
@@ -23,7 +22,10 @@ public SideloaderProcessing(HtmlDocumentProcessorPlugin htmlProcessorPlugin) {
         this.htmlProcessorPlugin = htmlProcessorPlugin;
     }
 
-    public ProcessedDocument processDocument(String url, String body, List<String> extraKeywords, int size) throws URISyntaxException {
+    public ProcessedDocument processDocument(String url,
+                                             String body,
+                                             List<String> extraKeywords,
+                                             int size) throws URISyntaxException {
         var crawledDoc = new CrawledDocument(
                 "encyclopedia.marginalia.nu",
                 url,

code/processes/converting-process/src/main/java/nu/marginalia/converting/sideload/stackexchange/StackexchangeSideloader.java

@@ -5,7 +5,7 @@
 import nu.marginalia.converting.sideload.SideloadSource;
 import nu.marginalia.integration.stackexchange.sqlite.StackExchangePostsDb;
 import nu.marginalia.keyword.DocumentKeywordExtractor;
-import nu.marginalia.language.sentence.SentenceExtractor;
+import nu.marginalia.language.sentence.ThreadLocalSentenceExtractorProvider;
 import nu.marginalia.model.EdgeDomain;
 import nu.marginalia.model.EdgeUrl;
 import nu.marginalia.model.crawl.DomainIndexingState;
@@ -29,20 +29,20 @@
 import java.util.concurrent.TimeUnit;
 
 public class StackexchangeSideloader implements SideloadSource {
-    private final SentenceExtractor sentenceExtractor;
+    private final ThreadLocalSentenceExtractorProvider sentenceExtractorProvider;
     private final DocumentKeywordExtractor keywordExtractor;
     private final String domainName;
 
     private final Path dbFile;
 
     @SneakyThrows
     public StackexchangeSideloader(Path pathToDbFile,
-                                   SentenceExtractor sentenceExtractor,
+                                   ThreadLocalSentenceExtractorProvider sentenceExtractorProvider,
                                    DocumentKeywordExtractor keywordExtractor
     ) {
         this.dbFile = pathToDbFile;
         this.domainName = StackExchangePostsDb.getDomainName(pathToDbFile);
-        this.sentenceExtractor = sentenceExtractor;
+        this.sentenceExtractorProvider = sentenceExtractorProvider;
         this.keywordExtractor = keywordExtractor;
     }
 
@@ -109,7 +109,7 @@ private ProcessedDocument convert(StackExchangePostsDb.CombinedPostModel post) {
 
             var url = new EdgeUrl(fullUrl);
             var doc = Jsoup.parse(fullHtml.toString());
-            var dld = sentenceExtractor.extractSentences(doc);
+            var dld = sentenceExtractorProvider.get().extractSentences(doc);
 
             ret.url = url;
             ret.words = keywordExtractor.extractKeywords(dld, url);