Community contribution from Alex Pelagenko improving our HTTP server

QuestDB is a high performance time-series database with SQL analytics that can power through data ingestion and analysis. It's open source and integrates with many tools and languages. Give us a try!

I have recently made a sizable contribution to QuestDB’s code and wanted to share my experience and feedback while it is still fresh in my head. I am not a complete outsider for the project and know Vlad personally but other than that it was voluntary to add a few lines of code to a project I like.

The HTTP server for QuestDB

QuestDB has a custom HTTP stack that uses non-blocking socket IO via a thin layer of JNI OS abstraction. Non-blocking IO is handled via two state machines. One for inbound traffic, which includes a series of parsing state machines. The other for outbound traffic. We focus on the outbound traffic state machine, which has to deal with two types of interruptions: slow socket on one side and data availability on the other. While slow socket interruption was already dealt with, the data availability interruption had been handled in a very trivial manner. When data was unavailable, the HTTP stack would report an immediate error and trigger a send-to-socket state machine.

Data availability interruptions are due to QuestDB’s single writer model. A table will be locked while the HTTP server is dealing with a CSV import request. A request to alter the locked table will bounce back with an error. Why is this interesting? It is a difficult problem of coordination amongst threads while at the same time keeping the whole stack non-blocking.

How I added queuing to QuestDB's HTTP stack

The first hurdle was to understand the stack, which is hard to follow at first glance. Control is passed around via both conditional statements and exception mechanisms. The thread messaging stack is also unusual. The API is non-blocking - the thread must find another task if the outbound queue is full or the inbound queue is empty.

Instead of rejecting requests due to data availability errors, I added a queuing system that catches the state of these requests in a priority queue. This queue is then processed by idle threads (idle because of IO interruptions) and retried at exponentially increasing intervals. For example the first retry will happen in 2ms then in 4ms, 8ms, 16ms … 512ms, 1s, 1s, 1s. The retry interval is a configuration parameter. Following this addition, operations are queued on the server, meaning that the user does not have to deal with errors or attempt to do this operation again. This piece of code processes priority queue:

private boolean sendToOutQueue() {
boolean useful = false;
final long now = clock.getTicks();
while (nextRerun.size() > 0) {
Retry next = nextRerun.peek();
if (next.getAttemptDetails().nextRunTimestamp <= now) {
useful = true;
Retry retry = nextRerun.poll();
if (!sendToOutQueue(retry)) {
return true;
else {
// All reruns are in the future.
return useful;
return useful;

QuestDB’s team was very patient explaining the different bits of the HTTP stack and guiding me to figure out how best to solve the multi writing problem. They were also careful not to dampen my enthusiasm with the pull request feedback.

Get the latest and greatest. Secure and never shared or sold.