Writing at work. Not writing here.

I haven't posted anything here in a long time. At the end of July I started a new job at a big company and have been busy learning how to be effective there. As with many big companies there is not much that I am allowed to share, especially when that company operates in a highly regulated industry. As I settle in and better understand the boundaries I am sure I will start writing here again. Not that anyone really cares; a blog is a vanity project after all.

Small set of guidelines for using Slack

I have been using Slack for a long time. Here are a few guidelines based on that experience:
  • Slack is part of your communications toolkit. It has limits. When you feel that you are typing too much or getting frustrated by misunderstandings then switch to a call. A consequence of this is that you should be prepared to make or accept a call — earbuds at the ready.
  • Slack is great for singular topic discussions. The channel clearly displays the discussion’s history. Even if the discussion spans many days the context is always available for a quick catch up.
  • Slack is terrible for overlapping topic discussions. The channel quickly becomes a hodgepodge of utterances and fragments that have little reference to the topical context. Overlapping topic channels are needed, however. For example, in one small company I worked at the #operations channel is primarily used to give notice of infrastructure changes. When using such a channel you need to remember that others will not have the same topical context as you. Where possible, include some reference to the topic — "Re X, we did Y." For example, "Re loud bang, we have flogged the responsible employee." Here "loud bang" is the strong reminder of the previous topical messages. See Threads. See Links Use pinned messages to define the purpose of the channel along with links to supporting materials. Messages are editable and so do improve and update the pinned message. Pinned messages can be quickly accessed via the "push pin" icon in the channel header.
  • Slack has discussion threads. Any message can be the start of a thread. Threads provide a natural grouping of related messages. The problem with them is that it is hard to have everyone on the channel use them consistently. For example, Jane started a thread to continue the discussion, but Jack, in haste, posted a message instead. Now we have 3 messages connected by two different mechanisms. Don’t use threads unless you can get everyone on the channel to use them consistently. (Actually, just don't use threads.)
  • When you want to clearly reference a previous message use a Slack link. It is long and ugly, but it is by far the most reliable reference.
  • We reuse text from many different applications and written languages everyday. It is common to copy from one place and paste it into Slack, and vise versa. When you do this you bring along a lot of unseen cruft. Cruft like formatting, odd visible characters, odd invisible characters, etc. And Slack does not help, in that it wants to pretty up the text for you; emoji interpretation being the most blatant example. When pasting into Slack always use Paste without formatting or Paste and Match style. And if you are pasting something technical like an email address, an access token, my salary, etc then use Slack's inline code formatting, eg "The email is `me@there.com`" becomes
  • Use the sidebar as a presence indicator. Making sure your immediate team members are favorited.

Macroservices, a middle ground between the monolith and microservices.

There is a middle ground between the monolith and microservices and that is "macroservices." Macroservices are distinguished by having one executable and many compositions. Each composition exercises a portion of the executable's internal components. Some of the components are exercised by all compositions (eg, identity management) and other components by singular composition (eg, a specialized data store). Composition deployments communicate with each other with REST or gRPC. Orchestration for resilience and scaling is accomplished using the same tools as for microservices.

A composition is nothing more than a declaration of a set of components to activate, their interdependencies, and their configurations. Configurations are generally properties, ie name and value pairs, accessed from the environment. Much as when you build the executable and draw all the benefits of type checking, unit testing, static analysis, etc you can build the composition and draw similar correctness assurances.

A macroservice allows your small team of developers to focus on what matters to your business, ie what your customers value. The development infrastructure is simple. The deployment infrastructure is flexible. Troubleshooting is comprehensible.

Written in response to Microservices are for companies with 500+ engineers.

Update: Perhaps a better name is "polyservices".

Update: I am reading Sam Newman's new book Monolith to Microservices: Evolutionary Patterns to Transform Your Monolith. My macroservice is more akin to his distributed monolith except that the services contained within do have stronger data independence than is typical of a monolith.

RI's Innovation Vouchers to help fund product development

I recently learned about RI's Innovation Voucher program. The vouchers are $5,000-$50,000 grants for small companies to buy the expertise needed to develop a new product or process. Vouchers can be redeemed for services at a research institution or to fund an in-house research & development project.

More information at RI Commerce Innovation Incentives.

Lucene, shadow query classes, and the visitor pattern

A good Lucene result is achieved from an index and a query working together.

Defining the schema for the index is mostly an upfront design task: What are the fields? What fields are stored and what are indexed? How are terms in fields parsed? Are found terms supplemented? Are multiple fields combined? Etc.

Once you have made these decisions revisiting them can be prohibitive without planning. Ie, reindexing is found to be too expensive in time or resources, or, worse, the source documents are no longer available. Most Lucene uses error on the side of having a cautious schema where a lot more is stored and indexed than is needed now with the hope of a successful schema refactoring in the future. My experience has been that you don't know enough now about your data, how to index it, how to query it, and how users want to access it to have a viable cautious schema. It is better to store all your sources so you can reindex when you know better how the indexes and queries can cooperate. This is just a cost to using a new technology (and, as always, it can be mitigated).

The upshot in the near term is that the index schema is static. You have to apply flexibility with the query.

As mentioned in a previous posting, there is a tendency to think of Lucene queries like SQL queries. That is, there is a single, correct rendition. Discard that thinking. There are no correct results; there are only better results. To achieve better results you need to watch what your users searches and workout how the queries need to adjust. For example, perhaps you discover that there is a shift in vocabulary happening. What was once "OS X" is now "macOS". When a user queries for OS X you need to also include macOS in the query.

The Lucene API contains a number of query subtypes. These are combined to construct an expression that characterizes the user's search intent. This nested data structure should be considered a starting point. This data structure will be augmented and reformed to reflect your current understanding of how to best use the indexes. In the example above, you want to include macOS when OS X is used.
The Lucene API query subclasses are too rigid for direct augmentation and reforming. In the past the API was downright unbending and so I developed a set of shadow query classes that were amenable for use with the Visitor pattern. For example, this shows two visitors, the first adds the macOS variant and the second converts the query to a Solr expression:

Query q = ... new TermQuery(10.0f, "f", "osx") ...
Map<String, List<String>> variants = new HashMap<>();
variants.put("a", Arrays.asList("osx, "macos"));
VariantsQueryVistor vistor = new VariantsQueryVistor(0.0001f, variants);
q = vistor.visitQuery(q);
s = new SolrLuceneQueryVistor().visitQuery(q).toString();

and s is

( "osx" OR "macos" ^ 0.0001 ) ^ 10.0

Git and the busy development shop

I don't like Git and this posting is not going to change anyone's mind about Git and I will continue to use Git as it has become the one version control system to rule them all.

Git does not work the way I want a version control tool to work and especially in a busy development shop. I work on multiple branches concurrently. Does anyone have serialized bug and feature work? I don't need the whole version control tree available all the time. I only need the trunk and the branches I am working on and I need them to be available simultaneously. Git's stash, which allows pausing development on one branch to work on another, is useless to me. Using it breaks my work flow and destabilizes my mental model of my file system. That is, when I am in ~/src/foo/branches/issue1/ I know everything below is work being done on issue1, and when I am in ~/src/foo/branches/feature2/ I know everything below is work being done on feature2. There is never any confusion as to what I am looking at. With Git, however, I need to frequently confirm that ~/src/foo/ is currently checked-out for issue1 or feature2. (Cloning to specific issue1 and feature2 directories is not a solution. See following.)

The other issue with Git is with centralized repos. Git's development model assumes many developers, each with their own full copy of the repo, and exchanging updates via patches. I understand this model and see its value. Introducing a centralized repo into this model adds complexity and collateral problems. It is not that the problems are unique to Git, but that a centralized Git repo intensives them. For example, developers A and B work on the same branch and each performs one or more checkins and some number of pushes to the origin (ie, the centralized repo). Since we have 3 repos in play -- A's repo, Bs repo, and the centralized repo -- the chances of a conflict after checkin are very high. Compare this with Subversion where there is only one repo. A and B can work as much as they like and will be stopped to resolve conflicts at checkin and not afterwards. I argue that resolving conflicts at checkin is a far better time to do this work as one's mind is in "differences" mode. When using Git I find myself performing a checkin immediately followed by a push so as to avoid this mess. Doing so loses a key feature of Git, ie distributed version control that works offline.

Update: When I wrote the following I was not aware of interactive rebase. An interactive rebase is used to reorder and combine commits so as to provide a list of changes where each change is functionally complete and the list is in a logical order. The video Interactive git rebase for code reviews and profit is a useful guide to doing that.

I hate Git's rebasing. File A is branched from master. A is changed N times on the branch and a few times on the master. Rebasing A is the equivalent of branching A from the current master and automatically applying the N changes. Doing this ruins the version history and reviewing the version history is the first step in fixing regressions. Other version history fraud comes from deleting branches and so loosing detail about who and when a change was introduced to the branch which are details that are often critical to recontextualizing the "fix". Git can be used without recourse to rebasing and deleting branches, but it seems to be common practice.

Back to coding...

Remain in control of your search

Lucene, Solr, and Elasticsearch have a powerful query language and a convenient textual representation. I have seen some API providers allow users to directly use this representation within their search API. Unfortunately, when you do so you loose significant control over your development and operations. You have exposed how you process your source data for full-text searching and so can no longer make behind-the-scenes schema changes. All your queryable data must be indexed in Lucene even if doing so is questionable and/or searchable elsewhere. You are locked into using Lucene to execute the query and at scale this can be exorbitantly expensive. To fix any of these issues you end up having to break backwards compatibility. How often will your API users accept this?

Instead, design your own query language. You can base it on Lucene's syntax if you like or an S-expression as it is trivial to parse in any programming language. No matter how you express the syntax, however, you have control over the semantics and the execution. For example, perhaps one of the searchable fields is not in your full-text index, but in a relational database and its presence in the query signals a programmatic join of the indexed and relational results. You can do that. Even if your queries are easily handled by Lucene at scale you are still better off translating yours syntax, like this old-school Google query

 -a +b c d

to the Lucene equivalent

 not a and ( b and (c d) )

because you remain in control.

Update: I fixed the term transposition in the Lucene equivalent of the Google expression. Doh!

Lucene boosting

Lucene has a powerful query language. The same language is used by Solr and Elasticsearch. I have found that users often don't utilize it well as they mistakenly apply their SQL experiences to it. With SQL your queries return exact results. Nothing in the result set is irrelevant. SQL query performance rewards tight queries consisting of few terms, few indexes, and few joins. With Lucene yours queries return ranked results. Much in the result set is actually peripheral. In fact, unless you limit the result set Lucene will return all the documents. A Lucene query is about getting a good ranking of results rather than exact results. To this end your query and indexes need to work together to achieve this. In general this means you need to make good use of broadening and narrowing terms, and boosting matches.

For example, if your query simply looked for the term "mouse" and you indexed your documents verbatim you should not expect to find any "mice". (Recall that Lucene sees your words as numbers and so "mouse" might be #23 and "mice" might be #6078.) It is therefore better to search for

mouse mice

When searching for "mouse mice" your results will be ordered so that either term gives equal weight to the document's rank. This is unlikely the correct course. The query was for "mouse" and you broadened it to include "mice". Documents matching "mouse" should be ranked higher than documents matching "mice". In Lucene's query language you do this by boosting the weight of terms. Documents matching "mouse" should be boosted orders of magnitude higher than "mice", ie

mouse^1000 mice

You will often see small boost values in other peoples examples. My experience has been that small boosts do not adequately differentiate documents. Big boosts do.

You likely noticed that my query was for "mouse mice" and not "mouse or mice". With Lucene, as with SQL, a boolean "or" kills performance. By not using "or" in your Lucene query you are allowing it to rank higher documents that contain both terms over documents that contain only one of them. Since the higher documents do contain the wanted term, "mouse", I don't see a need to repress ranking them equally to documents containing only "mouse" (ie, no use of "mice"). The performance cost is usually not worth it, especially when your queriers will be more complex than this simple example.

South Kingstown's impenetrable budget documents

The budgeting phrase "structural deficit" is a euphemism for spending more than you earn or spending less than necessary. That is, avoiding spending on maintenance or other costs that can be deferred or refinanced. The South Kingstown School District (SD) has clearly been running structural deficits. Implementing programs it can not afford. Delaying maintenance until unavoidable. And that this has been going on for years shows that the School Committee has failed us and the Town Council has inadequately performed their budgeting oversight.

The SD has a long history of providing impenetrable budget documents. Often the documents consist of a few printed spreadsheets and a mountain of slide decks. Both of which are incomprehensible without the oral narrative given at workshop meetings. Further, when you ask for the actual data you get files of account and values without any description of structure or purpose. This situation is unacceptable.

Not until the SD can provide an intelligible budget document that can be understood by us at our kitchen table should they be authorized to bond for any monies. Creating such a document is not a light undertaking, but luckily the SD can follow the exemplars found at Association of School Business Officials International (ASBO) and the Government Finance Officers Association (GFOA).

Authenticating S3 access using non-anonymous request URLs

If you run a small data center and have capped bandwidth you don't want to be delivering bulk data to customers. It is better to place the data in the cloud and redirect your customers to get the data there. Amazon's S3 is a good place for that as creating a public URL is trivial. If the data is not public then S3 has a simple mechanism for enabling you to authenticate access. To do this you run your own authentication service; this service prepares a signed, time limited URL that you give to the client to use to download the data from S3. The network interaction is all done within SSL and so you don't need to worry about the URL escaping into the wild and even if it did the loss is time limited.

The AWS S3 service calls this a non-anonymous request URL. For example, if your data is in the "2019-Q4.tsv" item in the "com.andrewgilmartin.bucket1" bucket the URL is


Your authentication service will (after authenticating the user) redirect the user's HTTP client to the URL


This is the non-anonymous request URL. The <<SIGNATURE>> is a base64 encoding of an SHA1 encryption of the HTTP method ("GET"), the path ("/com.andrewgilmartin.bucket1/2019-Q4.tsv"), and the expiration time (<<EXPIRES>>). The <<AWS_ACCESS_KEY>> corresponding secret key is used for the encryption. An example Java implementation is at S3RestAuthenticationUrlFactory.

For any of this to work you will need an AWS access key id and secret key that is associated with an IAM user with a policy to access the S3 bucket. If you have not done this before the video AWS S3 Bucket Security, Restrict Privileges to User using IAM Policy is a good tutorial. If you only want to allow read access then remove the "s3:PutObject" and "s3:DeleteObject" actions from the example policy.

Creating a Maven project for your web application

This posting continues the series on moving from an Ant to a Maven build.

The last stage is to actually move the Ant build to Maven. Your source tree is now quite spartan. It contains the web application, lots of configuration files, servlets or controllers, and non-core supporting classes. As before, you will create a new Maven project, establish dependencies, copy your files, and build and test until complete.

This project is a combination of webapp and Java but Maven does have an automated way of creating this. Instead, you need to first create the webapp project and then create the java tree. Create the webapp project

mvn archetype:generate \
  -DarchetypeGroupId=org.apache.maven.archetypes \
  -DarchetypeArtifactId=maven-archetype-webapp \
  -DarchetypeVersion=1.4 \
  -DinteractiveMode=false \
  -DgroupId=com.andrewgilmartin \
  -DartifactId=system-application \

Now create the Java tree

cd system-application
mkdir -p \
  src/main/java \
  src/main/resources \
  src/test/java \

The result is

├── pom.xml
└── src
    ├── main
    │   ├── java
    │   ├── resources
    │   └── webapp
    │       ├── WEB-INF
    │       │   └── web.xml
    │       └── index.jsp
    └── test
        ├── java
        └── resources

The pom.xml file is little different from those created before. The significant change is the <packaging/> element


The "war" value directs Maven to create the war instead of a jar (the default). Now add to the pom.xml the common and system-core dependencies, and any other dependencies specific to the application.

Your web application runs within a servlet container and that container provides some of your dependencies. You need these dependencies for compilation, but they should not be bundled into your war. Maven calls these "provided" dependencies. For these dependencies add a <scope/> element to your <dependency/> element, eg


Copy the application's code, configuration, and webapp from the system source tree to this project. Build and test as normal until you have clean results.

If you are interested in my help with your Ant to Maven transition contact me at andrew@andrewgilmartin.com.

Creating Maven projects for the core packages and command line tools

This posting continues the series on moving from an Ant to a Maven build.

With your common packages now having their own Maven build you can move on to the system itself. For this series I am assuming that your system is composed of a web application with several command line tools. The web application is likely a large set of servlets or Spring controllers. It's a monolith and it is going to stay that way for the near future. The command lines tools are used for nightly batch operations or ad hoc reports, etc. What they have in common is that they require some of the system's packages to function. Eg, they depend on its data access packages, protocol facilitation packages, billing logic packages, etc. The next stage is to separate the system's core code, the command line tools, and the application code and its configuration.

System Core Project

Create a new Maven project for the system core code

mvn archetype:generate \
  -DgroupId=com.andrewgilmartin \
  -DartifactId=system-core \
  -DarchetypeArtifactId=maven-archetype-quickstart \
  -DarchetypeVersion=1.4 \

Replace the groupId and artifactId as appropriate.

Copy all the system core code to this project much like you did when extracting the common code. You will likely again find that the core code has entanglements with non-core code that you are going to have to work out. That can be very difficult and require some refactoring; hopefully not significant enough to abandon the whole effort.

As you are assembling the system-core project you may discover that it tries to come to life. You have the Java equivalent of archaea and bacteria, ie a self configuring class or sets of classes. These are classes with static blocks, eg

public class Archaea {
    static { /* do some configuration */ }

That static block is executed as the class is used. Normally this has not been an issue as the classes were always used in the context of the whole system. Now they are isolated. If they depended on external resources or files that are no longer available then their initialization failures leave them in undefined states. You will need to work this out. Can the static block be eliminated or replaced with initialization upon first instance use? Maybe a Design Patterns refactoring is needed.

Build and test as normal until you have clean results.

Once your system-core project is complete remove its code from the system's source tree, remove unneeded dependencies from the Ant build.xml, and add the new dependency to the <mvn-dependencies/> element in build.xml. Build and test the system as normal until you have clean results.

Command Line Tool Projects

Now extract the command line tools from the system into their own Maven projects. These projects will depend on the system-common and system-core projects. The Maven build will also need to create an "uberjar", that is a single jar that bundles all the classes and jars needed to run the tool.

Pick a command line tool and create a new Maven project for it as you would normally. Eg, for the gizmo command line tool use

mvn archetype:generate \
  -DgroupId=com.andrewgilmartin \
  -DartifactId=gizmo \
  -DarchetypeArtifactId=maven-archetype-quickstart \
  -DarchetypeVersion=1.4 \

Replace the groupId and artifactId as appropriate. Add to the pom.xml the system-common and system-core dependencies, and any other dependencies specific to the tool. Copy the tool's code from the system source tree to this project. Build and test as normal until you have clean results.

To create the "uberjar" update pom.xml and replace the whole <plugins/> with


Replace "com.andrewgilmartin.gizmo.App" with the fully qualified class name of the tool. When you now build the Maven project you will see "maven-assembly-plugin" log

--- maven-assembly-plugin:3.1.0:single (assemble-all) @ gizmo ---
Building jar: /home/ajg/src/gizmo/target/gizmo-1.0-SNAPSHOT-jar-with-dependencies.jar

The file "gizmo-1.0-SNAPSHOT-jar-with-dependencies.jar" is the uberjar. To trial run your command line tool use

java -jar target/gizmo-1.0-SNAPSHOT-jar-with-dependencies.jar

Don't forget to add whatever command line options prevent the tool from doing any actual work!

Once your tool is complete remove its code from the system's source tree and remove unneeded dependencies from the Ant build.xml.

Continue this procedure for each of your command line tools.

Where are we

At this point you have

  1. System common code Maven project
  2. System core code Maven project
  3. Command line tools Maven projects
  4. Remaining system Ant project

The remaining system is just the web application with its configuration, servlets or controllers, and the odd ball classes that don't fit in system-common or system-core. The next stage is to refactor the system Ant project itself.

Some background on what I helped build and operate for many years at Crossref

For much of the last 10 years I have been working on the development and operations of a data management system. This system has two types of customers. The customers providing the data are publishers of science, technical, and medical journals, books, and organizers of conferences.

There are 14,000 publisher customers or their representatives. The customers using the data are also the publishers, but the primary users are researchers and those providing services to them. There are about 160,000 direct use customers. On average we make 300,000 changes per day to the data and service 8-15 million external requests per day.

The system is composed of statically allocated VMs running a common executable with each deployment having a function specific configuration. (All the VMs run CentOS 7.) We use several instances of HAProxy to dispatch external and internal requests to the appropriate deployment based on information in the URL and/or in HTTP headers. We run the same system in AWS and in our own data centers.

The common executable is a Java 8 implementation within a Spring 3 servlet framework. We build from a single code base and build a single WAR file for use in Tomcat 8. The Tomcat installations on the VMs have an identity and that identity is what the executable uses to configure itself on start up. Some of the deployments are microservices in that they manage their own data and are only accessible via network interface. Other deployments are distributed monoliths and so can be scaled out, but share Oracle and MySql replicated databases. All communication between deployments is over HTTP.

The application manages the canonical data in Oracle and MySql. (We only use MySql RDS in AWS.) Data is encoded in XML and deposited via HTTP API or administration webapp. All changes are made directly to the canonical data via JDBC. These changes are then propagated to secondary, function specific repositories on a periodic basis, ie the data is pushed. Secondary repositories are built upon MySql, AWS S3, Solr, Berkeley DB, and bespoke persistence data stores. Tertiary repositories are updated via change events using ActiveMQ (JMS). The different mechanisms for change propagation come from having to have firm control over consistency and stability for the secondary data and, less so, for the tertiary data. Operational logs are centralized and analyzed for deviations. Prometheus and bespoke ETL processes gather and present operational and business metrics.

The manager and I are the joint architects of the system. My responsibility is technical and his is line-of-business. We both architect operations, but I am not involved in hardware aspects of the data center.

There are 4 developers and 1 operations team members. (The manager is one of the 4 developers.) All developers do new development, maintenance development, and customer support. We use NetBeans for the IDE. We use Jira for issue tracking and light project management. We use Subversion for version control using feature branches and the trunk is always ready to release. We release every Tuesday. While the release process is efficient, we do not use a fully automated CI/CD process.

We do have an office location where 3 of the team members mostly work, but the manager and I are mostly remote. We also work closely with other staff in the UK and across the US. By necessity we operate as a remote-only team. We use Slack messaging, conference calls, and screen sharing.

The 24x7 production support is done by the operations member, the manager, and me.

I look forward to using these skills and learning new ones. I hope we can talk more about how I can help your team in the coming years.