Anais Dotis-Georgiou is a Developer Advocate for InfluxData with a passion for making data beautiful with the use of Data Analytics, AI, and Machine Learning. She takes the data that she collects, does a mix of research, exploration, and engineering to translate the data into something of function, value, and beauty. When she is not behind a screen, you can find her outside drawing, stretching, boarding, or chasing after a soccer ball.
InfluxData is the company building InfluxDB, the open source time series database used by more than a million developers around the world. Their mission is to help developers build intelligent, real-time systems with their time series data.
Can you share a bit about your journey from being a Research Assistant to becoming a Lead Developer Advocate at InfluxData? How has your background in data analytics and machine learning shaped your current role?
I earned my undergraduate degree in chemical engineering with a focus on biomedical engineering and eventually worked in labs performing vaccine development and prenatal autism detection. From there, I began programming liquid-handling robots and helping data scientists understand the parameters for anomaly detection, which made me more interested in programming.
I then became a sales development representative at Oracle and realized that I really needed to focus on coding. I took a coding boot camp at the University of Texas in data analytics and was able to break into tech, specifically developer relations.
I came from a technical background, so that helped shape my current role. Even though I didn’t have development experience, I could relate to and empathize with people who had an engineering background and mind but were also trying to learn software. So, when I created content or technical tutorials, I was able to help new users overcome technical challenges while placing the conversation in a context that was relevant and interesting to them.
Your work seems to blend creativity with technical expertise. How do you incorporate your passion for making data ‘beautiful’ into your daily work at InfluxData?
Lately, I’ve been more focused on data engineering than data analytics. While I don’t focus on data analytics as much as I used to, I still really enjoy math—I think math is beautiful, and will jump at an opportunity to explain the math behind an algorithm.
InfluxDB has been a cornerstone in the time series data space. How do you see the open source community influencing the development and evolution of InfluxDB?
InfluxData is very committed to the open data architecture and Apache ecosystem. Last year we announced InfluxDB 3.0, the new core for InfluxDB written in Rust and built with Apache Flight, DataFusion, Arrow, and Parquet–what we call the FDAP stack. As the engineers at InfluxData continue to contribute to those upstream projects, the community continues to grow and the Apache Arrow set of projects gets easier to use with more features and functionality, and wider interoperability.
What are some of the most exciting open-source projects or contributions you’ve seen recently in the context of time series data and AI?
It’s been cool to see the addition of LLMs being repurposed or applied to time series for zero-shot forecasting. Autolab has a collection of open time series language models, and TimeGPT is another great example.
Additionally, various open source stream processing libraries, including Bytewax and Mage.ai, that allow users to leverage and incorporate models from Hugging Face are pretty exciting.
How does InfluxData ensure its open source initiatives stay relevant and beneficial to the developer community, particularly with the rapid advancements in AI and machine learning?
InfluxData initiatives remain relevant and beneficial by focusing on contributing to open source projects that AI-specific companies also leverage. For example, every time InfluxDB contributes to Apache Arrow, Parquet, or DataFusion, it benefits every other AI tech and company that leverages it, including Apache Spark, DataBricks, Rapids.ai, Snowflake, BigQuery, HuggingFace, and more.
Time series language models are becoming increasingly vital in predictive analytics. Can you elaborate on how these models are transforming time series forecasting and anomaly detection?
Time series LMs outperform linear and statistical models while also providing zero-shot forecasting. This means you don’t need to train the model on your data before using it. There’s also no need to tune a statistical model, which requires deep expertise in time series statistics.
However, unlike natural language processing, the time series field lacks publicly accessible large-scale datasets. Most existing pre-trained models for time series are trained on small sample sizes, which contain only a few thousand—or maybe even hundreds—of samples. Although these benchmark datasets have been instrumental in the time series community’s progress, their limited sample sizes and lack of generality pose challenges for pre-training deep learning models.
That said, this is what I believe makes open source time series LMs hard to come by. Google’s TimesFM and IBM’s Tiny Time Mixers have been trained on massive datasets with hundreds of billions of data points. With TimesFM, for example, the pre-training process is done using Google Cloud TPU v3–256, which consists of 256 TPU cores with a total of 2 terabytes of memory. The pre-training process takes roughly ten days and results in a model with 1.2 billion parameters. The pre-trained model is then fine-tuned on specific downstream tasks and datasets using a lower learning rate and fewer epochs.
Hopefully, this transformation implies that more people can make accurate predictions without deep domain knowledge. However, it takes a lot of work to weigh the pros and cons of leveraging computationally expensive models like time series LMs from both a financial and environmental cost perspective.
This Hugging Face Blog post details another great example of time series forecasting.
What are the key advantages of using time series LMs over traditional methods, especially in terms of handling complex patterns and zero-shot performance?
The critical advantage is not having to train and retrain a model on your time series data. This hopefully eliminates the online machine learning problem of monitoring your model’s drift and triggering retraining, ideally eliminating the complexity of your forecasting pipeline.
You also don’t need to struggle to estimate the cross-series correlations or relationships for multivariate statistical models. Additional variance added by estimates often harms the resulting forecasts and can cause the model to learn spurious correlations.
Could you provide some practical examples of how models like Google’s TimesFM, IBM’s TinyTimeMixer, and AutoLab’s MOMENT have been implemented in real-world scenarios?
This is difficult to answer; since these models are in their relative infancy, little is known about how companies use them in real-world scenarios.
In your experience, what challenges do organizations typically face when integrating time series LMs into their existing data infrastructure, and how can they overcome them?
Time series LMs are so new that I don’t know the specific challenges organizations face. However, I imagine they’ll confront the same challenges faced when incorporating any GenAI model into your data pipeline. These challenges include:
- Data compatibility and integration issues: Time series LMs often require specific data formats, consistent timestamping, and regular intervals, but existing data infrastructure might include unstructured or inconsistent time series data spread across different systems, such as legacy databases, cloud storage, or real-time streams. To address this, teams should implement robust ETL (extract, transform, load) pipelines to preprocess, clean, and align time series data.
- Model scalability and performance: Time series LMs, especially deep learning models like transformers, can be resource-intensive, requiring significant compute and memory resources to process large volumes of time series data in real-time or near-real-time. This would require teams to deploy models on scalable platforms like Kubernetes or cloud-managed ML services, leverage GPU acceleration when needed, and utilize distributed processing frameworks like Dask or Ray to parallelize model inference.
- Interpretability and trustworthiness: Time series models, particularly complex LMs, can be seen as “black boxes,” making it hard to interpret predictions. This can be particularly problematic in regulated industries like finance or healthcare.
- Data privacy and security: Handling time series data often involves sensitive information, such as IoT sensor data or financial transaction data, so ensuring data security and compliance is critical when integrating LMs. Organizations must ensure data pipelines and models comply with best security practices, including encryption and access control, and deploy models within secure, isolated environments.
Looking forward, how do you envision the role of time series LMs evolving in the field of predictive analytics and AI? Are there any emerging trends or technologies that particularly excite you?
A possible next step in the evolution of time series LMs could be introducing tools that enable users to deploy, access, and use them more easily. Many of the time series LMs I’ve used require very specific environments and lack a breadth of tutorials and documentation. Ultimately, these projects are in their early stages, but it will be exciting to see how they evolve in the coming months and years.
Thank you for the great interview, readers who wish to learn more should visit InfluxData.
