Keeping your database up to date with Differential Dataflow

Dr. Roland Kuhn

Dr. Roland Kuhn

CTO and co-founder at Actyx

Given an audit trail of everything that has happened in your factory, how do you keep the dashboards, reports, and ERP system in sync with reality? This boils down to the problem of maintaining materialized views of the event log. The state of the art is to update the views — also in external systems — in a minimal fashion, changing only what needs to change and exactly when the events happen.


See also the blog post on for a higher-level overview.

The problem setting

In a factory, people and machines work together to produce goods, for example chairs. Multiple intermediate products like legs need to be manufactured before the chair can be assembled, where each piece is created and refined over a series of process steps at different workstations. Two obvious questions for shop-floor personnel are: What is currently going on, and how much time has it taken for the various people and machines to produce a given batch of chairs? These questions are merely examples of the ones we could answer if we had an audit trail like this:

8:52Fredstarts setting up Drill1 for order 4711
8:56Fredfinishes setup
8:56Drill1starts working on order 4711
8:56Fredstarts drilling holes for order 4711
9:12Fredreports 27 chair legs produced for order 4711
9:26Fredreports 13 more chair legs produced for order 4711
9:27Fredstops drilling holes for order 4711
9:27Drill1stops working on order 4711
9:27Fredreports hole drilling for order 4711 is finished

Computers can help obtain such an audit trail, many of the entries can even be created automatically or with very little additional input from a worker like Fred. One thing we need to ensure, though, is that this reporting does not keep Fred nor the drill from performing their duty. The IT system must be as reliable as paper — while being much easier to analyse later. This is why ActyxOS uses a fully decentralised approach, recording the events from the table above on the edge devices and synchronising between devices whenever a network connection is available.

What we want to see

In this post we consider two simple cases out of the many that should be implemented at the factory where Fred is working:

  • a shop-floor dashboard shall show what is currently going on at Drill1 (analog for other workstations)
  • a reporting database should be filled with summary information, one row per timespan that the machine was working for an order
  • the ERP system should receive a booking for how much time Fred and Drill1 have spent working on this production step of order 4711 once Fred says that it is finished

The first part can be implemented using Grafana if we keep a table in PostgreSQL up to date with one row containing the information of what is going on per workstation (e.g. which order is being processed and since when and by whom).

machinedoing whatsince
Drill1working on order 47118:56
Drill3working on order 47128:33

The second part can be implemented similarly, by adding a row to a table of timespans whenever a machine stops working on an order.


The third part works analog by creating an ERP transaction with the relevant bookings.

How we want to program it

In order to get these processes right, it would be best to describe the required database changes or ERP transactions based on patterns in the event log and then let a smart framework like Differential Dataflow figure out how to run the corresponding computations as well as what changes to commit when. Much of this work can be done using filtering and aggregation, like using SQL on a relational database.

let (injector, events) = Flow::<Event<MachineEvent>, _>::new(scope);
let latest = events
.map(|ev| match ev.payload {
MachineEvent::Started { order } => {
DashboardEntry::working(, order, ev.timestamp)
MachineEvent::Stopped { .. } => {
DashboardEntry::idle(, ev.timestamp)
.group_by(|entry| entry.machine.clone())
.max_by(|entry| entry.since)

This code snippet operates on a Flow of events, which is a DSL built on the differential-dataflow Rust library.

  • A flow is created within a scope of execution (see the full example for all details), returning an injector handle by which events can later be fed into this flow plus the events handle with which the data transformations are now described.
  • The .filter() method removes events which do not pertain to drills (just as an example), like WHERE in SQL,
  • the .map() turns each event into a machine status dashboard entry, like SELECT.

The resulting Flow at this point describes a collection of status records, one for each relevant event for each machine on the shop-floor. Since we are only interested in the most recent status update,

  • we use .group_by() to split the collection into one group per machine and then
  • take the maximum entry within each group sorted by timestamp.

The latest variable now holds a description of a collection that contains one record per machine. Whenever a new event is injected, this collection is updated accordingly. In the example, when the event from Drill1 comes in at 8:56 that the machine has started working on order 4711 the latest collection will have its previous record for workstation Drill1 removed and a new one inserted, denoting that Drill1 has been working on order 4711 since 8:56.

Going beyond SQL

One complication when working with an event log is that we need to correlate different records (events) to compute for example the duration column in the reporting table. This can be done in the dataflow DSL using the .reduce() combinator, which takes a function that turns a vector of inputs into a vector of outputs. We do this separately for each machine (again using .group_by()) to match each stop event with the corresponding start event for the same order, compute the duration between them and emit a machine usage record.

let records = events
.map(|ev| Excerpt {
lamport: ev.lamport,
event: ev.payload,
timestamp: ev.timestamp,
.group_by(|excerpt| excerpt.machine.clone())
.reduce(|_machine, inputs, outputs| {
let mut started_events = BTreeMap::new();
for (excerpt, _) in inputs {
// inputs are in ascending order, so we know that stop comes after start
match &excerpt.event {
MachineEvent::Started { order } => {
UsageEntry {
machine: excerpt.machine.clone(),
order: order.clone(),
started: excerpt.timestamp,
duration_micros: 0,
MachineEvent::Stopped { order } => {
if let Some(mut usage) = started_events.remove(&order) {
usage.duration_micros = excerpt.timestamp - usage.started;
outputs.push((usage, 1));

It is good practice to filter first and retain only the information necessary for later process steps because the .reduce() operator needs to keep all inputs in memory: when a new event is added for a machine, it will place it into its sorted slot in the inputs vector and run this function again. While this sounds wasteful, it has two important advantages:

  • it allows the dataflow framework to figure out what exactly changed in the outputs and only propagate that change further downstream (in our case: to the database)
  • and it allows the code to be written with full focus on the business logic, without distractions for state management

We’ll get back to the question of long-running inputs below.

To this end, it is often useful to introduce intermediate structures like Excerpt in the example above. Besides giving useful names to its fields, it also provides the necessary sort ordering so that the inputs are presented in causal order: we need to see effects after their cause, concretely we need to see the machine stop after it has started. Sorting by normal timestamp does not always achieve this, for example when the system clock of the edge device is modified or when there is clock skew between edge devices. Therefore, ActyxOS tracks causality using so-called Lamport clocks. The definition of the excerpt data structure is as follows:

#[derive(Debug, Clone, Ord, PartialOrd, Eq, PartialEq, Abomonation)]
struct Excerpt {
lamport: LamportTimestamp, // place first to sort in ascending (causal) order
machine: String,
event: MachineEvent,
timestamp: TimeStamp,

The derived trait instances are needed by the differential dataflow framework.

Writing to the database

So far we have described only the business logic, now we want to write the results into an actual database. First we need to spell out the details of such a record:

#[derive(Clone, Debug, Ord, PartialOrd, Eq, PartialEq, Abomonation)]
pub struct UsageEntry {
pub machine: String,
pub order: String,
pub started: TimeStamp, // microseconds since the Unix epoch
pub duration_micros: i64,

This record needs to be turned into one database row with four columns, to be inserted into a table of some name. This is explained to the database driver by implementing the DbRecord trait.

impl DbRecord<SqliteDbMechanics> for UsageEntry {
fn table_version() -> i32 {
fn table_name() -> &'static str {
fn columns() -> &'static [actyxos_data_flow::db::DbColumn] {
static X: &[DbColumn] = &[
DbColumn {
name: "machine",
tpe: "text not null",
exclude: false,
index: true,
DbColumn {
name: "manufacturing_order",
tpe: "text",
exclude: false,
index: false,
DbColumn {
name: "started",
tpe: "timestamp with time zone",
exclude: false,
index: true,
DbColumn {
name: "duration_micros",
tpe: "bigint",
exclude: false,
index: false,
fn values(&self) -> Vec<<SqliteDbMechanics as DbMechanics>::SqlValue> {
Box::new(self.started.as_i64() / 1_000_000),

This implementation is specific to the kind of database we want to write to because the column types may depend on this information and the Rust data type for column values depends on the database driver. In this case we’re targeting Sqlite3 because it doesn’t require setup. You’ll want to switch to PostgreSQL or Microsoft SQL Server for feeding your dashboards and reports.

Transactional storage

We have described the contents of the database so far in terms of declarative business logic and a table schema for our records. The final ingredient for knowing exactly which rows need to be in the database is to denote the set of input events that have been processed so far. For this reason, the database driver stores not only the records in the data table, it also stores — within the same transaction — the offset map of the events that have been ingested into an adjacent table. In the example above that table would be named usage_offsets.

This allows the process to be stopped and restarted without any loss of data: the business logic will be (re)run on all events whose records are not yet in the database. Storing data and offsets in the same transaction ensures that there can be neither duplicates nor losses.

Another benefit of this approach is that the restart of the exporter computes only the minimal amount of events needed to become operational again, making this operation complete much quicker than reprocessing all events would take.

Reading the events

The final part is how to get the events from the ActyxOS Event Service. Since we want to read the events in a Rust program, we first need to define their format and deserialization using serde.

#[derive(Clone, Debug, Ord, PartialOrd, Eq, PartialEq, Deserialize, Abomonation)]
#[serde(tag = "type", rename_all = "camelCase")]
pub enum MachineEvent {
Started { order: String },
Stopped { order: String },

This matches the events emitted by the sample webview app, whose Typescript definition is

export type Event =
| { type: 'started'; order: string }
| { type: 'stopped'; order: string }

With this, we have all the pieces to write the exporter’s main() function:

let mut db = SqliteDB::<Union<_>>::new("", "db_name")?;
let subscriptions = vec![Subscription::wildcard(semantics!("machineFish"))];
runtime.handle().clone(), // Tokio runtime to use for running async tasks
&mut db, // DB to store results in
"dashboard", // name for logging
move |offsets, to_db| {
subscriptions, // which events we need
offsets, // where we left off last time
to_db, // sending channel towards DB
"dashboard", // name for logging
1_000, // events per transaction

Besides the helper functions that wire together the EventService client, the differential dataflow machine, and the database driver, the main part here is the definition of the event subscriptions. This exporter needs to see all events emitted by the machineFish. The full code is available here.

Avoiding endless growth

The setup described above should typically be sufficient for more than a year of data collection and export. The limiting factor is the amount of state that needs to be stored within the dataflow pipelines to generate the correct deltas. Due to the short-lived nature of processes on the factory shop-floor, there is a natural solution to this problem: as e.g. manufacturing orders have a lifetime of hours, days, or a few weeks, we can be pretty sure that events older than a year will not have an influence on currently generated reporting data.

With this, we can provide a solution that avoids endless growth by restarting the exporter once per year, initializing the internal state only from the events of the year before. This gives enough context to the differential dataflow engine to emit the right deltas going forward. The actyxos_data_flow library supports this with the new_limited function to construct input collections:

let one_year = Duration::from_secs(365 * 86400);
let (injector, events) = Flow::<Event<MachineEvent>, _>::new_limited(scope, one_year);

This also avoids reading the whole event history upon first start, it will only ingest the last year of data — this can be very helpful in case the data collection has been ongoing in that factory for a much longer time already.


To summarize, actyxos_data_flow is a library that makes live data export from ActyxOS into transactional databases easy, efficient, and resilient. The process can be restarted at any time without data losses or duplications, and the programmer can concentrate fully on the business logic and the table schema without having to worry about how to get the events or how to keep track of what was already processed.