Analyses a secondary stream of information against a known point-of-truth and reports inconsistencies.

When you have a Lucene-based index of substantial size, say many hundreds of millions of records, what you want is confidence that your index is correct. In many cases, people use Solr/ElasticSearch/Compass to index their central database, mongodb, hbase etc so the index is a secondary storage of data.

How do you know if your index is accurate? Can you just reindex 500 million documents anytime you like? (That's the Aliens: "Nuke the site from Orbit... It's the only way to be sure" approach). No, if there ARE inconsistencies in your index, then you want to:

• find the items that are incorrect (and only them)
• do it fast

Scrutineer has been designed with this in mind, it can find any inconsistencies in your index fast.

Scrutineer relies on your data having 2 core properties:

• an ID - a unique identifier for your object
• a Version - something stored in your primary datastore for that object that represents the temporal state of that object

The Version property is commonly used in an Optimistic Locking pattern. If you store the ID & Version information in your secondary store (say, Solr/ElasticSearch) then you can always compare for any given item whether the version in secondary store is up to date.

Scrutineer takes a stream from your primary, and a stream from your secondary store, presumes they are sorted identically (more on that later) and walks the streams doing a merge comparison. It detects 4 states:

1. Both items are identical (yay!)
2. An ID is missing from the secondary stream (A missed add? maybe that index message you sent to Solr/ElasticSearch never made it, anyway, it's not there)
3. An ID was detected in the secondary, but wasn't in the primary stream (A missed delete? something was deleted on the primary, but the secondary never got the memo)
4. The ID exists in both streams, but the Version values are inconsistent (A missed update? similar to the missed add, this time perhaps an update to a row in your DB never made it to Solr/ElasticSearch)

Here's an example, 2 streams in sorted order, one from the Database (your point-of-truth), and one from ElasticSearch (the one you're checking) with the : for each side:

Scrutineer picks up that:

• ID '2' is missing from ElasticSearch
• ID '5' was deleted from the database at some point, but ElasticSearch still has it in there
• ID '6' is visible in ElasticSearch but appears to have the wrong version

The very first thing you'll need to do is get your JDBC Driver jar and place it in the 'lib' directory of the unpacked package. We already have a JTDS driver in there if you're using SQL Server (that's just what we use).

bin/scrutineer \
            --jdbcURL=jdbc:jtds:sqlserver://mydbhost/mydb  \
            --jdbcDriverClass=net.sourceforge.jtds.jdbc.Driver \
            --jdbcUser=itasecret \
            --jdbcPassword=itsasecret   \
            --sql="select id,version from myobjecttype order by cast(id as varchar(100))" \
            --clusterName=mycluster \
            --indexName=myindex \
            --query="_type:myobjecttype" \
            --numeric

Note: if you're weirded out about that '...cast(...)' then don't worry, we'll explain that shortly.

• jdbcURL - Standard JDBC URL you would use for your app to connect to your database
• jdbcDriverClass - Fully qualified class name of your JDBC Driver (don't forget to put your JDBC Driver jar in the lib directory as said above!)
• jdbcUser - User account to access your JDBC Database
• jdbcPassword - Password required for the user credentials
• sql - The SQL used to generate a lexicographical stream of ID & Version values (in that column order)
• clusterName - This is your ElasticSearch cluster name used to autodetect and connect to a node in your cluster
• indexName - The name of the index on your ElasticSearch cluster
• query - A query_parser compatible search query that returns all documents in your ElasticSearch index relating to the SQL query you're using. Since it is common for an index to contain a type-per-db-table you can use the "_type:" search query to filter for all values for that type.
• numeric - Use this if your query returns results numerically ordered

Scrutineer writes any inconsistencies direct to Standard Error, in a well-defined, tab-separated format for easy parsing to feed into a system to reindex/cleanup. If we use the Example scenario above, this is what Scrutineer would print out:

NOTINSECONDARY    2    23455
MISMATCH    6    34666    secondaryVersion=34556
NOTINPRIMARY    5    38475

The general format is:

FailureType\tID\tVERSION\tOptional:Additional Info

This means you are missing this item in your secondary and you should reindex/re-add to your secondary stream

This means the version of the object stored in the secondary is not the same information as the primary, and you should reindex

The object was removed from the Primary store, but the secondary still has it. You should remove this item from your secondary.

Scrutineer does not report when items match, we'll presume you're just fine with that...

If you're connecting to a remote host, you need to specify the host name:

--hostName=my.elastic.search.net

By default the version is taken from the _version meta-field, but you can specify a custom field contained in the document source as well:

--versionField=my_version_field

If you use timestamps as your version property, it is sometimes useful to know the underlying time value of the timestamp to triage why the differences are occurcing. For example, if a MISMATCH error occurs, you can look at the timestamp version of the Primary to work out when the last update was done to see why the Secondary never received it, perhaps by digging through your own application logs.

By default, Scrutineer doesn't presume this, so you can use the

--versions-as-timestamps

Command-line option to tell Scrutineer that your versions are timestamps, and a slightly different formatting of the result is chosen. Each time the Version property is reported (in both Primary or Secondary values, if printed) also include a human-readable, ISO8601 timestamp using the servers local Timezone for convenience.

For example, using the above output sample:

NOTINSECONDARY    2    23455(1970-1-1T16:30:55.000+10:00)
MISMATCH    6    34666(1970-1-1T19:37:46.000+10:00)    secondaryVersion=34556(1970-1-1T19:35:56.000+10:00)
NOTINPRIMARY    5    38475(1970-1-1T20:41:15.000+10:00)

The TAB value is still the field delimiter here.

By default, Scrutineer allocates 256m to the Java Heap, which is used for sort, and ElasticSearch result buffers. This should be more than enough for the majority of cases but if you find you get an OutOfMemoryError, you can override the JAVA_OPTS environment variable to provide more heap. e.g.

export JAVA_OPTS=-Xmx1048m

VERY IMPORTANT: Scrutineer relies on both streams to be sorted using an identical mechanism. It requires input streams to be in lexicographical (default) or numerical (indicate using --numeric) sort order.

Since Aconex uses ElasticSearch, Scrutineer supports ES out of the box, but it would not be difficult for others to integrate a Solr stream and wire something up. Happy to take Pull Requests!

The authors of Scrutineer, Aconex, index content from a JDBC data source and index using ElasticSearch. We do the following:

• In the database table of the object being indexed we add an Insert/Update trigger to populate a 'lastupdated' timestamp column as our Version property
• When we index into ElasticSearch, we set the Version property of the item using the VersionType.EXTERNAL setting.
• We create an SQL Index on this tuple so these 2 fields can be retrieved from the database very fast

• Your Version property is Long compatible. You can use java.sqlTimestamps column types too as a Version (that's what we do)
• Aconex is DB->ElasticSearch centric at the moment. We've tried to keep things loosely coupled, so it should be simple to add further integration points for other Primary & Secondary sources (HBase, MongoDB, Solr).

Scrutineer ships with the SQL Server JTDS driver by default (it's what we use). All you should need to do is drop your own JDBC driver in the 'repo' sub-directory of the Scrutineer distribution (where all the other jars are). We use the Maven AppAssembler plugin which is configured to automatically load all JARs in this path onto the classpath.

Scrutineer is a Maven project, which really should just build right out of the box if you have Maven installed. Just type:

mvn package

And you should have a Tarball in the 'target' sub-directory.

First, Please add unit tests!

Second, Please add integration tests!

Third, We have tightened up the quality rule set for CheckStyle, PMD etc pretty hard. Before you issue a pull request, please run:

mvn verify

which will run all quality checks. Sorry to be super-anal, but we just like Clean Code.

• Scrutineer currently only runs in a single thread based on a single stream.
  It would be good to provide a 'manifest' to Scrutineer to outline a set of stream verifications to perform, perhaps one for each type you have so that your multi-core system can perform multiple stream comparisons in parallel.

• Incremental checking – Right now Scrutineer checks the whole thing, BUT if you are using timestamp-based versions, there's no reason it couldn't only check objects that were changed after the last known full verification. This would require one to keep track of deletes on the primary stream (perhaps an OnDelete Trigger in your SQL database) so that IDs that were deleted in the primary stream after the last full check could be detected correctly.

• Obviously we'd love to have a Solr implementation here, we hope the community can help here.