2 files changed, 18 insertions, 3 deletions
diff --git a/scalding/src/main/scala/sandcrawler/CdxBackfillJob.scala b/scalding/src/main/scala/sandcrawler/CdxBackfillJob.scala
index f98b6e9..389a96a 100644
--- a/scalding/src/main/scala/sandcrawler/CdxBackfillJob.scala
+++ b/scalding/src/main/scala/sandcrawler/CdxBackfillJob.scala
@@ -3,6 +3,7 @@ package sandcrawler
 import java.util.Properties
 
 import scala.util.Try
+import scala.util.matching.Regex
 import scala.util.parsing.json.JSONObject
 
 import cascading.pipe.joiner._
@@ -122,11 +123,14 @@ object CdxBackfillJob {
   }
 
   def keepCdx(line: CdxLine) : Boolean = {
+    val sha1Pattern = """[A-Z2-7]{32}""".r
     if (List(line.surt, line.datetime, line.url, line.mime, line.c_size, line.offset, line.warc).contains("-")) {
-      // TODO: hadoop counter (was: "DASHLINE")
       false
-    } else if (line.httpStatus != "200" || line.sha1.size != 32) {
-      // TODO: sha1.isalnum()
+    } else if (line.httpStatus != "200") {
+      false
+    } else if (line.mime != "application/pdf") {
+      false
+    } else if (sha1Pattern.unapplySeq(line.sha1).isEmpty) {
       false
     } else if (List(line.c_size, line.offset, line.datetime).map(s => Try(s.toLong).toOption).contains(None)) {
       false
diff --git a/scalding/src/test/scala/sandcrawler/CdxBackfillJob.scala b/scalding/src/test/scala/sandcrawler/CdxBackfillJob.scala
index a6107fc..c092f7f 100644
--- a/scalding/src/test/scala/sandcrawler/CdxBackfillJob.scala
+++ b/scalding/src/test/scala/sandcrawler/CdxBackfillJob.scala
@@ -35,6 +35,17 @@ class CdxBackfillTest extends FlatSpec with Matchers {
     // redirect
     assert(false === keepCdx(lineToCdxLine(
       "edu,upenn,ldc)/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf 20170828233154 https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf application/pdf 301 WL3FEA62TEU4F52Y5DOVQ62VET4QJW7G - - 210251 931661233 SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828231135742-00000-00009-wbgrp-svc284/SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828232253025-00005-3480~wbgrp-svc284.us.archive.org~8443.warc.gz")))
+    // not PDF
+    assert(false === keepCdx(lineToCdxLine(
+      """edu,upenn,ldc)/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf 20170828233154 https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf text/plain 200 WL3FEA62TEU4F52Y5DOVQ62VET4QJW7G - - 210251 931661233 SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828231135742-00000-00009-wbgrp-svc284/SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828232253025-00005-3480~wbgrp-svc284.us.archive.org~8443.warc.gz""")))
+    // invalid base32 SHA1
+    assert(false === keepCdx(lineToCdxLine(
+      """edu,upenn,ldc)/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf 20170828233154 https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf application/pdf 200 WL3FE010101010101010101VET4QJW7G - - 210251 931661233 SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828231135742-00000-00009-wbgrp-svc284/SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828232253025-00005-3480~wbgrp-svc284.us.archive.org~8443.warc.gz""")))
+    assert(false === keepCdx(lineToCdxLine(
+      """edu,upenn,ldc)/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf 20170828233154 https://www.ldc.upenn.edu/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf application/pdf 200 WL33FEA62TEU4F52Y5DOVQ62VET4QJW7G - - 210251 931661233 SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828231135742-00000-00009-wbgrp-svc284/SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828232253025-00005-3480~wbgrp-svc284.us.archive.org~8443.warc.gz""")))
+    // dashed field
+    assert(false === keepCdx(lineToCdxLine(
+      """edu,upenn,ldc)/sites/www.ldc.upenn.edu/files/medar2009-large-arabic-broadcast-collection.pdf 20170828233154 - application/pdf 200 WL3FEA62TEU4F52Y5DOVQ62VET4QJW7G - - 210251 931661233 SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828231135742-00000-00009-wbgrp-svc284/SEMSCHOLAR-PDF-CRAWL-2017-08-04-20170828232253025-00005-3480~wbgrp-svc284.us.archive.org~8443.warc.gz""")))
   }
 
   it should "know what CDX lines are" in {