mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-29 07:34:31 +00:00
84 lines
2.3 KiB
TeX
84 lines
2.3 KiB
TeX
\section{Production Deployment}
|
|
|
|
\label{sec:production}
|
|
|
|
\subsection{Infrastructure}
|
|
|
|
Hanzo Flow is deployed as part of the Hanzo Platform:
|
|
|
|
\begin{itemize}[leftmargin=1.1em]
|
|
\item \textbf{Editor}: React application with canvas-based graph editor, running on CDN edge nodes.
|
|
\item \textbf{Execution engine}: Go service on Kubernetes (4 replicas, auto-scaling to 16), processing workflow executions.
|
|
\item \textbf{Trace store}: PostgreSQL + S3 for execution traces (compressed, 90-day retention).
|
|
\item \textbf{LLM access}: Via Hanzo LLM Gateway with per-workflow cost budgets.
|
|
\end{itemize}
|
|
|
|
\subsection{Usage Statistics (10 Months)}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\begin{tabular}{lr}
|
|
\toprule
|
|
\textbf{Metric} & \textbf{Value} \\
|
|
\midrule
|
|
Total workflows created & 1,247 \\
|
|
Total executions & 8,412,893 \\
|
|
Avg. nodes per workflow & 9.2 \\
|
|
Avg. execution time & 12.4s \\
|
|
End-to-end success rate & 94.1\% \\
|
|
Unique users & 3,218 \\
|
|
LLM calls processed & 34,891,247 \\
|
|
Total cost (LLM + compute) & \$847,231 \\
|
|
Avg. cost per execution & \$0.10 \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Production statistics (May 2025 -- Feb 2026).}
|
|
\label{tab:prod_stats}
|
|
\end{table}
|
|
|
|
\subsection{Pattern Distribution}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\small
|
|
\begin{tabular}{lcc}
|
|
\toprule
|
|
\textbf{Pattern} & \textbf{Workflows} & \textbf{Success Rate} \\
|
|
\midrule
|
|
Sequential chain & 31.2\% & 97.3\% \\
|
|
Router & 18.4\% & 95.8\% \\
|
|
Parallel ensemble & 14.7\% & 92.1\% \\
|
|
Iterative refinement & 12.3\% & 89.4\% \\
|
|
Map-reduce & 10.1\% & 93.7\% \\
|
|
Human-in-the-loop & 8.2\% & 96.2\% \\
|
|
Recursive decomposition & 5.1\% & 84.3\% \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Workflow pattern distribution in production.}
|
|
\label{tab:patterns}
|
|
\end{table}
|
|
|
|
\subsection{Common Failure Modes}
|
|
|
|
Analysis of failed executions reveals:
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\small
|
|
\begin{tabular}{lcc}
|
|
\toprule
|
|
\textbf{Failure Mode} & \textbf{Frequency} & \textbf{Recovery} \\
|
|
\midrule
|
|
LLM format error & 34\% & Prompt retry \\
|
|
API timeout & 21\% & Backoff retry \\
|
|
Model rate limit & 18\% & Model fallback \\
|
|
Type mismatch (runtime) & 12\% & Manual fix \\
|
|
Cost budget exceeded & 9\% & User notification \\
|
|
Infinite loop detected & 6\% & Auto-terminate \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Failure mode analysis of production executions.}
|
|
\label{tab:failures}
|
|
\end{table}
|
|
|