IWMI notified me that AReS was down with an HTTP 502 error
Looking at UptimeRobot I see it has been down for 33 hours, but I never got a notification
I don’t see anything in the Elasticsearch container logs, or the systemd journal on the host, but I notice that the angular_nginx container isn’t running
IWMI notified me that AReS was down with an HTTP 502 error
Looking at UptimeRobot I see it has been down for 33 hours, but I never got a notification
I don’t see anything in the Elasticsearch container logs, or the systemd journal on the host, but I notice that the angular_nginx container isn’t running
<li>IWMI notified me that AReS was down with an HTTP 502 error
<ul>
<li>Looking at UptimeRobot I see it has been down for 33 hours, but I never got a notification</li>
<li>I don’t see anything in the Elasticsearch container logs, or the systemd journal on the host, but I notice that the <code>angular_nginx</code> container isn’t running</li>
<li>I simply started it and AReS was running again:</li>
<li>Margarita from CCAFS emailed me to say that workflow alerts haven’t been working lately
<ul>
<li>I guess this is related to the SMTP issues last week</li>
<li>I had fixed the config, but didn’t restart Tomcat so DSpace didn’t load the new variables</li>
<li>I ran all system updates on CGSpace (linode18) and DSpace Test (linode26) and rebooted the servers</li>
</ul>
</li>
</ul>
<h2id="2021-06-03">2021-06-03</h2>
<ul>
<li>Meeting with AMCOW and IWMI to discuss AMCOW getting IWMI’s content into the new AMCOW Knowledge Hub
<ul>
<li>At first we spent some time talking about DSpace communities/collections and the REST API, but then they said they actually prefer to send queries to sites on the fly and cache them in Redis for some time</li>
<li>That’s when I thought they could perhaps use the OpenSearch, but I can’t remember if it’s possible to limit by community, or only collection…</li>
<li>Looking now, I see there is a “scope” parameter that can be used for community or collection, for example:</li>
<li>Looking at the Solr statistics on CGSpace for last month I see many requests from hosts using seemingly normal Windows browser user agents, but using the MSN bot’s DNS
<ul>
<li>For example, user agent <code>Mozilla/5.0 (compatible; MSIE 9.0; Windows NT 6.0; Trident/5.0; Trident/5.0)</code> with DNS <code>msnbot-131-253-25-91.search.msn.com.</code></li>
<li>I queried Solr for all hits using the MSN bot DNS (<code>dns:*msnbot* AND dns:*.msn.com.</code>) and found 457,706</li>
<li>I extracted their IPs using Solr’s CSV format and ran them through my <code>resolve-addresses.py</code> script and found that they all belong to MICROSOFT-CORP-MSN-AS-BLOCK (AS8075)</li>
<li>Note that <ahref="https://www.bing.com/webmasters/help/how-to-verify-bingbot-3905dc26">Microsoft’s docs say that reverse lookups on Bingbot IPs will always have “search.msn.com”</a> so it is safe to purge these as non-human traffic</li>
<li>I purged the hits with <code>ilri/check-spider-ip-hits.sh</code> (though I had to do it in 3 batches because I forgot to increase the <code>facet.limit</code> so I was only getting them 100 at a time)</li>
</ul>
</li>
<li>Moayad sent a pull request a few days ago to re-work the harvesting on OpenRXV
<ul>
<li>It will hopefully also fix the duplicate and missing items issues</li>
<li>I had a Skype with him to discuss</li>
<li>I got it running on podman-compose, but I had to fix the storage permissions on the Elasticsearch volume after the first time it tries (and fails) to run:</li>
<li>Then I uploaded the resulting CSV to CGSpace, updating 161 items</li>
<li>Start a harvest on AReS</li>
<li>I found <ahref="https://jira.lyrasis.org/browse/DS-1977">a bug</a> and <ahref="https://github.com/DSpace/DSpace/pull/2584">a patch</a> for the private items showing up in the DSpace sitemap bug
<ul>
<li>The fix is super simple, I should try to apply it</li>
</ul>
</li>
</ul>
<h2id="2021-06-21">2021-06-21</h2>
<ul>
<li>The AReS harvesting finished, but the indexes got messed up again</li>
<li>I was looking at the JSON export I made yesterday and trying to understand the situation with duplicates
<ul>
<li>We have 90,000+ items, but only 85,000 unique:</li>
<li>Some appear twice in the Elasticsearch index, but appear in only one collection</li>
<li>Some appear twice in the Elasticsearch index, and appear in <em>two</em> collections</li>
<li>Some appear twice in the Elasticsearch index, but appear in three collections (!)</li>
</ul>
</li>
<li>So really we need to just check whether a handle exists before we insert it</li>
<li>I tested the <ahref="https://github.com/DSpace/DSpace/pull/2584">pull request for DS-1977</a> that adjusts the sitemap generation code to exclude private items
<ul>
<li>It applies cleanly and seems to work, but we don’t actually have any private items</li>
<li>The issue we are having with AReS hitting restricted items in the sitemap is that the items have restricted metadata, not that they are private</li>
</ul>
</li>
<li>Testing the <ahref="https://github.com/DSpace/DSpace/pull/2275">pull request for DS-4065</a> where the REST API’s <code>/rest/items</code> endpoint is not aware of private items and returns an incorrect number of items
<ul>
<li>This is most easily seen by setting a low limit in <code>/rest/items</code>, making one of the items private, and requesting items again with the same limit</li>
<li>I confirmed the issue on the current DSpace 6 Demo:</li>
<li>I tested the pull request on DSpace Test and it works, so I left a note on GitHub and Jira</li>
<li>Last week I noticed that the Gender Platform website is using “cgspace.cgiar.org” links for CGSpace, instead of handles
<ul>
<li>I emailed Fabio and Marianne to ask them to please use the Handle links</li>
</ul>
</li>
<li>I tested the <ahref="https://github.com/DSpace/DSpace/pull/2543">pull request for DS-4271</a> where Discovery filters of type “contains” don’t work as expected when the user’s search term has spaces
<ul>
<li>I tested with filter “farmer managed irrigation systems” on DSpace Test</li>
<li>Before the patch I got 293 results, and the few I checked didn’t have the expected metadata value</li>
<li>After the patch I got 162 results, and all the items I checked had the exact metadata value I was expecting</li>
<li>Make a <ahref="https://github.com/atmire/COUNTER-Robots/pull/43">pull request</a> to the COUNTER-Robots project to add two new user agents: crusty and newspaper
<ul>
<li>These two bots have made ~3,000 requests on CGSpace</li>
<li>Then I added them to our local bot override in CGSpace (until the above pull request is merged) and ran my bot checking script:</li>
<li>These bots account for ~42,000 hits in our statistics… I will just purge them and add them to our local override, but I can’t be bothered to submit them to COUNTER-Robots since I’d have to look up the information for each one</li>
<li>I re-synced DSpace Test (linode26) with the assetstore, Solr statistics, and database from CGSpace (linode18)</li>
</ul>
<h2id="2021-06-23">2021-06-23</h2>
<ul>
<li>I woke up this morning to find CGSpace down
<ul>
<li>The logs show a high number of abandoned PostgreSQL connections and locks:</li>
</ul>
</li>
</ul>
<pre><codeclass="language-console"data-lang="console"># journalctl --since=today -u tomcat7 | grep -c 'Connection has been abandoned'
978
$ psql -c 'SELECT * FROM pg_locks pl LEFT JOIN pg_stat_activity psa ON pl.pid = psa.pid;' | wc -l
10100
</code></pre><ul>
<li>I sent a message to Atmire, hoping that the database logging stuff they put in place last time this happened will be of help now</li>
<li>In the mean time, I decided to upgrade Tomcat from 7.0.107 to 7.0.109, and the PostgreSQL JDBC driver from 42.2.20 to 42.2.22 (first on DSpace Test)</li>
<li>I also applied the following patches from the 6.4 milestone to our <code>6_x-prod</code> branch:
<ul>
<li>DS-4065: resource policy aware REST API hibernate queries</li>
<li>DS-4271: Replaced brackets with double quotes in SolrServiceImpl</li>
</ul>
</li>
<li>After upgrading and restarting Tomcat the database connections and locks were back down to normal levels:</li>
</ul>
<pre><codeclass="language-console"data-lang="console">$ psql -c 'SELECT * FROM pg_locks pl LEFT JOIN pg_stat_activity psa ON pl.pid = psa.pid;' | wc -l
63
</code></pre><ul>
<li>Looking in the DSpace log, the first “pool empty” message I saw this morning was at 4AM:</li>
</ul>
<pre><codeclass="language-console"data-lang="console">2021-06-23 04:01:14,596 ERROR org.hibernate.engine.jdbc.spi.SqlExceptionHelper @ [http-bio-127.0.0.1-8443-exec-4323] Timeout: Pool empty. Unable to fetch a connection in 5 seconds, none available[size:250; busy:250; idle:0; lastwait:5000].
</code></pre><ul>
<li>Oh, and I notice 8,000 hits from a Flipboard bot using this user-agent:</li>
</ul>
<pre><codeclass="language-console"data-lang="console">Mozilla/5.0 (Macintosh; Intel Mac OS X 10.11; rv:49.0) Gecko/20100101 Firefox/49.0 (FlipboardProxy/1.2; +http://flipboard.com/browserproxy)
</code></pre><ul>
<li>We can purge them, as this is not user traffic: <ahref="https://about.flipboard.com/browserproxy/">https://about.flipboard.com/browserproxy/</a>
<ul>
<li>I will add it to our local user agent pattern file and eventually submit a pull request to COUNTER-Robots</li>
</ul>
</li>
<li>I merged <ahref="https://github.com/ilri/OpenRXV/pull/96">Moayad’s health check pull request in AReS</a> and I will deploy it on the production server soon</li>
</ul>
<h2id="2021-06-24">2021-06-24</h2>
<ul>
<li>I deployed the new OpenRXV code on CGSpace but I’m having problems with the indexing, something about missing the mappings on the <code>openrxv-items-temp</code> index
<ul>
<li>I extracted the mappings from my local instance using <code>elasticdump</code> and after putting them on CGSpace I was able to harvest…</li>
<li>But still, there are way too many duplicates and I’m not sure what the actual number of items should be</li>
<li>According to the OAI ListRecords for each of our repositories, we should have about:
<ul>
<li>MELSpace: 9537</li>
<li>WorldFish: 4483</li>
<li>CGSpace: 91305</li>
<li>Total: 105325</li>
</ul>
</li>
<li>Looking at the last backup I have from harvesting before these changes we have 104,000 total handles, but only 99186 unique:</li>
<li>So the harvest on the live site is missing items, then why didn’t the add missing items plugin find them?!
<ul>
<li>I notice that we are missing the <code>type</code> in the metadata structure config for each repository on the production site, and we are using <code>type</code> for item type in the actual schema… so maybe there is a conflict there</li>
<li>I will rename type to <code>item_type</code> and add it back to the metadata structure</li>
<li>The add missing items definitely checks this field…</li>
<li>I modified my local backup to add <code>type: item</code> and uploaded it to the temp index on production</li>
<li>Oh! nginx is blocking OpenRXV’s attempt to read the sitemap:</li>
<li>I fixed nginx so it always allows people to get the sitemap and then re-ran the plugins… now it’s checking 180,000+ handles to see if they are collections or items…
<ul>
<li>I see it fetched the sitemap three times, we need to make sure it’s only doing it once for each repository</li>
</ul>
</li>
<li>According to the api logs we will be adding 5,697 items:</li>
<li>Spent a few hours with Moayad troubleshooting and improving OpenRXV
<ul>
<li>We found a bug in the harvesting code that can occur when you are harvesting DSpace 5 and DSpace 6 instances, as DSpace 5 uses numeric (long) IDs, and DSpace 6 uses UUIDs</li>
<li>We can apparently get the names of the jobs in each hash using <code>hget</code>:</li>
</ul>
<pre><codeclass="language-console"data-lang="console">127.0.0.1:6379> TYPE bull:plugins:401827
hash
127.0.0.1:6379> HGET bull:plugins:401827 name
"dspace_add_missing_items"
</code></pre><ul>
<li>I whipped up a one liner to get the keys for all plugin jobs, convert to redis <code>HGET</code> commands to extract the value of the name field, and then sort them by their counts:</li>
| sed -e 's/^bull/HGET bull/' -e 's/\([[:digit:]]\)$/\1 name/' \
| ncat -w 3 localhost 6379 \
| grep -v -E '^\$' | sort | uniq -c | sort -h
3 dspace_health_check
4 -ERR wrong number of arguments for 'hget' command
12 mel_downloads_and_views
129 dspace_altmetrics
932 dspace_downloads_and_views
186428 dspace_add_missing_items
</code></pre><ul>
<li>Note that this uses <code>ncat</code> to send commands directly to redis all at once instead of one at a time (<code>netcat</code> didn’t work here, as it doesn’t know when our input is finished and never quits)
<ul>
<li>I thought of using <code>redis-cli --pipe</code> but then you have to construct the commands in the redis protocol format with the number of args and length of each command</li>
</ul>
</li>
<li>There is clearly something wrong with the new DSpace health check plugin, as it creates WAY too many jobs every time we run the plugins</li>
<li>Looking into the spike in PostgreSQL connections last week
<ul>
<li>I see the same things that I always see (large number of connections waiting for lock, large number of threads, high CPU usage, etc), but I also see almost 10,000 DSpace sessions on 2021-06-25</li>
<li>Annoyingly I found 37,000 more hits from Bing using <code>dns:*msnbot* AND dns:*.msn.com.</code> as a Solr filter
<ul>
<li>WTF, they are using a normal user agent: <code>Mozilla/5.0 (Windows NT 6.3; Trident/7.0; rv:11.0) like Gecko</code></li>
<li>I will purge the IPs and add this user agent to the nginx config so that we can rate limit it</li>
</ul>
</li>
<li>I signed up for Bing Webmaster Tools and verified cgspace.cgiar.org with the BingSiteAuth.xml file
<ul>
<li>Also I adjusted the nginx config to explicitly allow access to <code>robots.txt</code> even when bots are rate limited</li>
<li>Also I found that Bing was auto discovering all our RSS and Atom feeds as “sitemaps” so I deleted 750 of them and submitted the real sitemap</li>
<li>I need to see if I can adjust the nginx config further to map the <code>bot</code> user agent to DNS like msnbot…</li>
</ul>
</li>
<li>Review Abdullah’s filter on click pull request
<ul>
<li>I rebased his code on the latest master branch and tested adding filter on click to the map and list components, and it works fine</li>
<li>There seems to be a bug that breaks scrolling on the page though…</li>
<li>Abdullah fixed the bug in the filter on click branch</li>
<li>Integrate <code>prettier</code> into the frontend and backend and format everything on the <code>master</code> branch</li>
<li>Re-work the GitHub Actions workflow for frontend and add one for backend</li>
<li>The workflows run <code>npm install</code> to test dependencies, and <code>npm ci</code> with <code>prettier</code> to check formatting</li>
<li>Also I merged Abdallah’s filter on click pull request</li>
</ul>
</li>
</ul>
<h2id="2021-06-30">2021-06-30</h2>
<ul>
<li>CGSpace is showing a blank white page…
<ul>
<li>The status is HTTP 200, but it’s blank white… so UptimeRobot didn’t send a notification!</li>
</ul>
</li>
<li>The DSpace log shows:</li>
</ul>
<pre><codeclass="language-console"data-lang="console">2021-06-30 08:19:15,874 ERROR org.hibernate.engine.jdbc.spi.SqlExceptionHelper @ Cannot get a connection, pool error Timeout waiting for idle object
</code></pre><ul>
<li>The first one of these I see is from last night at 2021-06-29 at 10:47 PM</li>
<li>I restarted Tomcat 7 and CGSpace came back up…</li>
<li>I didn’t see that Atmire had responded last week (on 2021-06-23) about the issues we had
<ul>
<li>He said they had to do the same thing that they did last time: switch to the postgres user and kill all activity</li>
<li>He said they found tons of connections to the REST API, like 3-4 per second, and asked if that was normal</li>
<li>I pointed him to our Tomcat server.xml configuration, saying that we purposefully isolated the Tomcat connection pools between the API and XMLUI for this purpose…</li>
<li>Export a list of all CGSpace’s AGROVOC keywords with counts for Enrico and Elizabeth Arnaud to discuss with AGROVOC:</li>
</ul>
<pre><codeclass="language-console"data-lang="console">localhost/dspace63= > \COPY (SELECT DISTINCT text_value AS "dcterms.subject", count(*) FROM metadatavalue WHERE dspace_object_id in (SELECT dspace_object_id FROM item) AND metadata_field_id = 187 GROUP BY "dcterms.subject" ORDER BY count DESC) to /tmp/2021-06-30-agrovoc.csv WITH CSV HEADER;
COPY 20780
</code></pre><ul>
<li>Actually Enrico wanted NON AGROVOC, so I extracted all the center and CRP subjects (ignoring system office and themes):</li>
</ul>
<pre><codeclass="language-console"data-lang="console">localhost/dspace63= > \COPY (SELECT DISTINCT LOWER(text_value) AS subject, count(*) FROM metadatavalue WHERE dspace_object_id in (SELECT dspace_object_id FROM item) AND metadata_field_id IN (119, 120, 127, 122, 128, 125, 135, 203, 208, 210, 215, 123, 236, 242) GROUP BY subject ORDER BY count DESC) to /tmp/2021-06-30-non-agrovoc.csv WITH CSV HEADER;
COPY 1710
</code></pre><ul>
<li>Fix an issue in the Ansible infrastructure playbooks for the DSpace role
<ul>
<li>It was causing the template module to fail when setting up the npm environment</li>
<li>We needed to install <code>acl</code> so that Ansible can use <code>setfacl</code> on the target file before becoming an unprivileged user</li>
</ul>
</li>
<li>I saw a strange message in the Tomcat 7 journal on DSpace Test (linode26):</li>
</ul>
<pre><codeclass="language-console"data-lang="console">Jun 30 16:00:09 linode26 tomcat7[30294]: WARNING: Creation of SecureRandom instance for session ID generation using [SHA1PRNG] took [111,733] milliseconds.
</code></pre><ul>
<li>What’s even crazier is that it is twice that on CGSpace (linode18)!</li>
<li>Apparently OpenJDK defaults to using <code>/dev/random</code> (see <code>/etc/java-8-openjdk/security/java.security</code>):</li>
<li><code>/dev/random</code> blocks and can take a long time to get entropy, and urandom on modern Linux is a cryptographically secure pseudorandom number generator
<ul>
<li>Now Tomcat starts much faster and no warning is printed so I’m going to add this to our Ansible infrastructure playbooks</li>
</ul>
</li>
<li>Interesting resource about the lore behind the <code>/dev/./urandom</code> workaround that is posted all over the Internet, apparently due to a bug in early JVMs: <ahref="https://bugs.java.com/bugdatabase/view_bug.do?bug_id=6202721">https://bugs.java.com/bugdatabase/view_bug.do?bug_id=6202721</a></li>
<li>I’m experimenting with using PgBouncer for pooling instead of Tomcat’s JDBC</li>
