merge backfill into extraction directory

2 files changed, 74 insertions, 0 deletions

diff --git a/extraction/tests/files/example.cdx b/extraction/tests/files/example.cdx
new file mode 100644
index 0000000..84e3271
--- /dev/null
+++ b/extraction/tests/files/example.cdx
@@ -0,0 +1,20 @@
+edu,cmu,cs,adm,reports-archive)/anon/usr0/ftp/usr0/anon/2002/cmu-cs-02-119.pdf 20170706005950 http://reports-archive.adm.cs.cmu.edu/anon/usr0/ftp/usr0/anon/2002/CMU-CS-02-119.pdf application/pdf 200 MQHD36X5MNZPWFNMD5LFOYZSFGCHUN3V - - 361006 17120058 CITESEERX-CRAWL-2017-06-20-20170706004100259-00924-00932-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170706005946792-00926-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+fi,tkk,lib)/diss/2001/isbn951225459x/isbn951225459x.pdf 20170705074926 http://lib.tkk.fi/Diss/2001/isbn951225459X/isbn951225459X.pdf application/pdf 200 KJBCOT7LGBNIAVGEGPUELK5OK6RTFORR - - 344175 255650124 CITESEERX-CRAWL-2017-06-20-20170705074433815-00129-00138-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705074843696-00134-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,oxfordjournals,nar)/cgi/reprint/gkl1060v1.pdf 20170706035441 http://nar.oxfordjournals.org/cgi/reprint/gkl1060v1.pdf text/html 301 OX6MLVDFURLT2KSYCXUYW2PZNOVFSEVF - - 697 49346051 CITESEERX-CRAWL-2017-06-20-20170706034741172-00140-00149-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706035435634-00148-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+org,ifaamas)/proceedings/aamas09/pdf/01_full%20papers/02_08_fp_0272.pdf 20170706081902 http://www.ifaamas.org/Proceedings/aamas09/pdf/01_Full%20Papers/02_08_FP_0272.pdf application/pdf 200 GYHX35QJWRJELWJ5GDQZPTPOUUZOCTKF - - 251180 34635154 CITESEERX-CRAWL-2017-06-20-20170706081825105-00419-00428-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706081838210-00420-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+de,fau,cs)/publications/2014/lukas_14_masterthesis.pdf 20170705101722 http://www4.cs.fau.de/Publications/2014/lukas_14_masterthesis.pdf application/pdf 200 GIUQT7SXZ33TWEFBM2MWURJI2M3QE3IW - - 1290532 71068435 CITESEERX-CRAWL-2017-06-20-20170705101605019-00279-00288-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705101714659-00281-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+de,bund,jki,pub)/index.php/jabfq/article/download/3568/4462 20170706041152 http://pub.jki.bund.de/index.php/JABFQ/article/download/3568/4462/ text/html 301 XZBNO24W2ZPQQMJYE6YUUCSRUF7G3ZBT - - 552 417292708 CITESEERX-CRAWL-2017-06-20-20170706040506112-00160-00169-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706041021844-00165-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+whois://whois.arin.net/z+%2B+132.177.133.114 20170713120653 whois://whois.arin.net/z+%2B+132.177.133.114 text/plain - IDEID4YQ6MVJSOE57NPVDLL53ZB3J4DX - - 876 30983517 CITESEERX-CRAWL-2017-06-20-20170707064626094-01007-01015-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170711214025652-01014-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+za,co,csir,researchspace)/dspace/bitstream/10204/4048/1/smith2_2010.pdf 20170706094159 http://researchspace.csir.co.za/dspace/bitstream/10204/4048/1/Smith2_2010.pdf unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 446 104830407 CITESEERX-CRAWL-2017-06-20-20170706093829986-00509-00518-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706094137978-00512-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+org,annals)/article.aspx?articleid=705034 20170707013120 http://annals.org/article.aspx?articleid=705034 text/html 301 QQYKL57QSERLFM3LXSWMNOFXMOCN7C5G - - 22665 28113974 CITESEERX-CRAWL-2017-06-20-20170707013100780-00967-00976-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170707013100780-00967-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+org,annals)/pdfaccess.ashx?url=/data/journals/aim/20105/0000605-200512200-00013.pdf 20170707045304 http://annals.org/pdfaccess.ashx?url=/data/journals/aim/20105/0000605-200512200-00013.pdf text/html 302 423S7EMGLCVIZ3FLVD7TLAG75HWE4RGI - - 644 222908628 CITESEERX-CRAWL-2017-06-20-20170707042504366-00997-01006-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170707045044604-00999-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+com,sagepub,spi)/content/28/4/501.full.pdf 20170705092027 http://spi.sagepub.com/content/28/4/501.full.pdf unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 396 553180242 CITESEERX-CRAWL-2017-06-20-20170705091311851-00219-00228-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705091759818-00223-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+ir,mediaj)/favicon.ico 20170705075240 http://mediaj.ir/favicon.ico text/html 404 E3WSNQ7JAFOW7N3ZJ6GLV27T52T25JDK - - 589 455827180 CITESEERX-CRAWL-2017-06-20-20170705074433815-00129-00138-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705075051100-00135-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+com,sagepub,jpr)/content/8/3-4/239.full.pdf 20170705074931 http://jpr.sagepub.com/content/8/3-4/239.full.pdf unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 400 270368088 CITESEERX-CRAWL-2017-06-20-20170705074433815-00129-00138-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705074843696-00134-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+jp,co,nittuden)/business/pdf/transparent_thermoplastic_resin_with_electron_beam_cross-linking.pdf 20170706083459 http://www.nittuden.co.jp/business/pdf/Transparent_Thermoplastic_Resin_with_Electron_Beam_Cross-Linking.pdf application/pdf 200 V32E3CCO7NMI2M4OHLKG73DXD72LR4B2 - - 715081 761088410 CITESEERX-CRAWL-2017-06-20-20170706082646066-00429-00438-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706083257353-00436-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+lt,lms)/robots.txt 20170705122708 http://www.lms.lt/robots.txt text/plain 200 PF3HTQQT2ULYRWFLJGUWZKHTVZUVMZ2F - - 592 668333707 CITESEERX-CRAWL-2017-06-20-20170705121748408-00399-00408-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705122352502-00406-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+hu,bme,phy)/~szalay/pub/multipartcriteriaposter.pdf 20170705124828 http://www.phy.bme.hu/%7Eszalay/pub/multipartcriteriaPoster.pdf application/pdf 200 L3TUEEZLBJTHAVH74B5N426FAIDBCCOE - - 187866 964760782 CITESEERX-CRAWL-2017-06-20-20170705123641979-00419-00428-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705124315591-00426-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,adb,openaccess)/bitstream/handle/11540/1260/new-regime-sme-finance-asia.pdf;jsessionid=f966a3bdac9882ec5a7c326b130f6f81?sequence=1 20170705090940 https://openaccess.adb.org/bitstream/handle/11540/1260/new-regime-sme-finance-asia.pdf%3Bjsessionid%3DF966A3BDAC9882EC5A7C326B130F6F81?sequence%3D1 unk 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 515 634039376 CITESEERX-CRAWL-2017-06-20-20170705090333400-00209-00218-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705090728803-00212-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,physiology,ajpregu)/content/272/4/r1084 20170706131006 http://ajpregu.physiology.org/content/272/4/R1084 text/html 200 3FOQSKT4WBYOUA6VKKJCEQCN6QF35ANT - - 27346 336293585 CITESEERX-CRAWL-2017-06-20-20170706130432396-00707-00716-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706130850866-00711-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
+de,desy,www-it)/common/documentation/cd-docs/sc2002/paperpdf/pap234.pdf 20170705121813 http://www-it.desy.de/common/documentation/cd-docs/SC2002/paperpdf/pap234.pdf application/pdf 200 BONCZ4NNGRNYR22ASFVU7VYTQ24RRNP4 - - 72421 381715704 CITESEERX-CRAWL-2017-06-20-20170705120827801-00389-00398-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705121708700-00397-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+org,oxfordjournals,bmb)/content/28/3/247.full.pdf 20170706014948 http://bmb.oxfordjournals.org/content/28/3/247.full.pdf text/html 301 EJWYVOPONJRARK7SGG6COFRN7CSTHROY - - 643 119398161 CITESEERX-CRAWL-2017-06-20-20170706014800946-00020-00029-wbgrp-svc285/CITESEERX-CRAWL-2017-06-20-20170706014907678-00022-3671~wbgrp-svc285.us.archive.org~8443.warc.gz
diff --git a/extraction/tests/test_backfill_hbase_from_cdx.py b/extraction/tests/test_backfill_hbase_from_cdx.py
new file mode 100644
index 0000000..1a13e5b
--- /dev/null
+++ b/extraction/tests/test_backfill_hbase_from_cdx.py
@@ -0,0 +1,54 @@
+"""
+TODO: could probably refactor to use unittest.mock.patch('happybase')
+"""
+
+import io
+import json
+import pytest
+import mrjob
+import happybase_mock
+from backfill_hbase_from_cdx import MRCDXBackfillHBase
+
+@pytest.fixture
+def job():
+    """
+    Note: this mock only seems to work with job.run_mapper(), not job.run();
+    the later results in a separate instantiation without the mock?
+    """
+    conn = happybase_mock.Connection()
+    conn.create_table('wbgrp-journal-extract-test',
+        {'file': {}, 'grobid0': {}, 'f': {}})
+    table = conn.table('wbgrp-journal-extract-test')
+
+    job = MRCDXBackfillHBase(['--no-conf', '-'], hb_table=table)
+    return job
+
+
+def test_some_lines(job):
+
+    raw = io.BytesIO(b"""
+com,sagepub,cep)/content/28/9/960.full.pdf 20170705062200 http://cep.sagepub.com/content/28/9/960.full.pdf application/pdf 301 3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ - - 401 313356621 CITESEERX-CRAWL-2017-06-20-20170705061647307-00039-00048-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705062052659-00043-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+eu,eui,cadmus)/bitstream/handle/1814/36635/rscas_2015_03.pdf;jsessionid=761393014319a39f40d32ae3eb3a853f?sequence=1 20170705062202 http://cadmus.eui.eu/bitstream/handle/1814/36635/RSCAS_2015_03.pdf%3Bjsessionid%3D761393014319A39F40D32AE3EB3A853F?sequence%3D1 application/PDF 200 MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J - - 854156 328850624 CITESEERX-CRAWL-2017-06-20-20170705061647307-00039-00048-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705062052659-00043-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+com,pbworks,educ333b)/robots.txt 20170705063311 http://educ333b.pbworks.com/robots.txt text/plain 200 6VAUYENMOU2SK2OWNRPDD6WTQTECGZAD - - 638 398190140 CITESEERX-CRAWL-2017-06-20-20170705062707827-00049-00058-wbgrp-svc284/CITESEERX-CRAWL-2017-06-20-20170705063158203-00053-31209~wbgrp-svc284.us.archive.org~8443.warc.gz
+""")
+
+    job.sandbox(stdin=raw)
+    job.run_mapper()
+
+    assert job.hb_table.row(b'1') == {}
+    # HTTP 301
+    assert job.hb_table.row(b'sha1:3I42H3S6NNFQ2MSVX7XZKYAYSCX5QBYJ') == {}
+    # valid
+    assert job.hb_table.row(b'sha1:MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J') != {}
+    # text/plain
+    assert job.hb_table.row(b'sha1:6VAUYENMOU2SK2OWNRPDD6WTQTECGZAD') == {}
+
+    row = job.hb_table.row(b'sha1:MPCXVWMUTRUGFP36SLPHKDLY6NGU4S3J')
+    assert row[b'file:mime'] == b"application/pdf"
+
+    file_cdx = json.loads(row[b'file:cdx'].decode('utf-8'))
+    assert int(file_cdx['offset']) == 328850624
+
+    f_c = json.loads(row[b'f:c'].decode('utf-8'))
+    assert f_c['u'] == "http://cadmus.eui.eu/bitstream/handle/1814/36635/RSCAS_2015_03.pdf%3Bjsessionid%3D761393014319A39F40D32AE3EB3A853F?sequence%3D1"
+    assert b'i' not in f_c