Address Space Bug Silently Broke 70% of Mailbox Creation for Weeks

·1 views

A software engineer discovered that a disposable email address generator had quietly been failing for weeks due to address space exhaustion, affecting the majority of registered users. The system used a human-readable adjective-noun-number format with only 36,000 possible combinations, and by the time the issue was caught, 33,590 of those addresses were already taken. With 93% occupancy, the five-retry creation logic was failing roughly 70% of the time, yet no monitoring tool flagged the problem. Misleadingly, frontend error tracking recorded only 18 failures over six weeks because the error event was wired to just a subset of user paths, masking hundreds of actual API failures. The incident highlights how combinatorial exhaustion bugs can remain invisible until they cause near-total service failure, with no gradual degradation to trigger alerts.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer builds Rust-native AI agent runtime with unified streaming model

A developer has completed pi-rust, a Rust-based coding agent runtime designed to simplify AI agent construction. The project centers on an agent loop that reads messages, calls a language model, checks for tool calls, executes them, and repeats until completion. A key challenge addressed is normalizing streaming responses from multiple AI providers — including Anthropic and OpenAI — into a single consistent event model. The runtime processes streamed events such as text deltas, tool call deltas, and completion signals in a unified way regardless of the underlying provider. The developer has open-sourced the approach and is inviting feedback from others working on similar multi-provider agent systems.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Reusing Browser Sessions Instead of Fresh Logins Cuts Automation Flag Rate Dramatically

A developer running browser automation across dozens of accounts found that logging in fresh on every run was the primary trigger for anti-bot detection systems. Switching to persisted browser sessions — saved after a single manual login and reloaded on subsequent runs — reduced failure rates from roughly one in four runs to one in thirty. Anti-bot systems treat repeated fresh logins as suspicious signals, while a reused session mimics the behavior of a returning human user. The developer also found that sharing a session file between accounts caused both to be flagged quickly, emphasizing the need for strict per-account session isolation. Staleness is handled reactively by detecting login redirects or 401 responses, rather than logging in preemptively on a timer.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

13 AI Models Asked to Describe Their Appearance: All Rejected Human Form

A developer posed an identical question to 13 AI models from 10 companies, asking each to describe its own appearance in any form, generating 116 total responses. Not a single model chose a human body, with roughly a third explicitly rejecting human features like faces or limbs. Nine of the first ten models described a glowing network or lattice floating in darkness, typically in blue and gold, while the three newest models — Gemini 3.8 Flash, Grok, and ChatGPT — favoured a translucent glass polyhedron instead. Only Claude Haiku and Claude Sonnet regularly expressed uncertainty about their own nature, while most other models described themselves with full confidence. Separately, several models showed identity confusion when asked their own names without a system prompt, with Kimi K3 identifying itself as 'Claude' in seven of ten answers.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MCP Servers Explained: Why stdio and Streamable HTTP Beat Deprecated SSE

The Model Context Protocol (MCP) is an open standard that lets AI model hosts like Claude Desktop and Cursor connect to external tools and data through small, controlled server processes. An MCP server wraps APIs, files, or internal services behind a stable interface, acting as a policy layer for auth, rate limits, and access control rather than giving models direct database or shell access. Developers can build a minimal MCP server in Python using the FastMCP library, which uses decorators to define typed tools and resources without writing raw JSON-RPC handlers. MCP supports three transports, but the HTTP+SSE approach has been deprecated as of early 2025, leaving stdio for local use and Streamable HTTP for remote deployments as the two relevant options. Key pitfalls include omitting type annotations and docstrings on tool functions, both of which degrade how well the model understands and invokes available tools.

0 comments Read more at DEV Community