- Delay DeviceCache initial load by 1 second to allow Redis connections
- Move shutdown handlers to end of supervision tree
- This prevents shutdown handlers from interfering with startup failures
Co-Authored-By: Claude <noreply@anthropic.com>
- Add SKIP_DB_CREATE env var to skip database creation when using PgBouncer
- Fix redis_children() to return proper child specs
- Implement zero-downtime graceful shutdown without user notifications
- Add health check plug for Kubernetes probes
- Remove user-facing shutdown notifications for seamless deployments
Co-Authored-By: Claude <noreply@anthropic.com>
This commit introduces a dual-mode caching and rate limiting system that can use
either Redis (for distributed deployments) or ETS/Cachex (for single-node setups).
Key changes:
- Add Cache abstraction layer that automatically switches between Redis and Cachex
- Implement RedisCache module with full distributed cache functionality
- Implement RedisRateLimiter with sliding window algorithm for accurate rate limiting
- Add RateLimiterWrapper to provide unified API for both implementations
- Update application startup to conditionally use Redis when REDIS_URL is set
- Migrate all cache operations to use the new Cache abstraction
- Support graceful fallback to ETS-based solutions when Redis is unavailable
The system automatically detects Redis availability via REDIS_URL environment
variable and switches between implementations without code changes. This enables
proper distributed caching and rate limiting in Kubernetes deployments while
maintaining backward compatibility for development environments.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
The phoenix_pubsub_redis adapter expects the url option at the top level,
not nested under redis_options.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Remove static Redis PubSub configuration from runtime.exs
- Add dynamic pubsub_config/0 function in application.ex
- Configure PubSub adapter based on cluster_enabled and REDIS_URL at startup
- Ensures proper initialization order for distributed deployments
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Both /status and /status.json endpoints now show cluster-wide APRS-IS
connection status instead of just local node status.
Features:
- get_cluster_aprs_status() checks all nodes in cluster via RPC
- Returns status from any connected node (usually the leader)
- Shows cluster info: total nodes, connected nodes, leader node
- Graceful fallback to local status in non-clustered mode
- Enhanced status page UI with cluster information display
This ensures status pages accurately reflect the true APRS-IS
connection state across the entire cluster.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
The cleanup_stale_registrations function was calling Process.alive?
on remote PIDs, which causes an ArgumentError. Fixed by:
- Using Process.alive? only for local PIDs
- Using :rpc.call for remote PID liveness checks
- Adding proper error handling for RPC failures
- Cleaning up registrations for disconnected nodes
This prevents GenServer crashes when checking leader status.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Leader election now waits for other nodes to connect before starting
- Added cleanup of stale global registrations from disconnected nodes
- Only starts election after cluster is established in clustered mode
- Non-clustered mode continues to elect immediately
This prevents both nodes from becoming leaders when they start
before the cluster is fully formed.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
The previous conflict resolution tried to call Process.info/2 on remote
PIDs, which doesn't work. Changed to use node name comparison for
deterministic leader selection across the cluster.
This ensures only one leader is elected when nodes join the cluster.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Prevent both AprsIsConnection and cluster-managed Is connection from
starting simultaneously. When clustering is enabled, only the cluster's
ConnectionManager should handle APRS-IS connections via leader election.
This resolves aprsc server disconnections caused by duplicate validated
clients with the same username (W5ISP-1).
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Implement leader election to ensure only one APRS-IS connection across
multiple Kubernetes replicas. This prevents duplicate packet processing
and respects APRS-IS usage policies.
Key changes:
- Add leader election using :global registry
- Create connection manager for dynamic APRS-IS management
- Implement packet distribution from leader to all nodes
- Add Kubernetes headless service for node discovery
- Configure DNS-based clustering with libcluster
- Update deployment to support 3 replicas with clustering
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Replace get_in/Access usage with proper struct pattern matching to fix
UndefinedFunctionError when accessing Sentry.Event fields. The struct
doesn't implement the Access behavior, so we now use pattern matching
and direct field access instead.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add create_database/0 function to handle database creation before migrations
- Make DeviceCache handle missing tables gracefully on startup
- Retry device loading after 5 seconds if initial load fails
- Prevent deployment failures when database doesn't exist yet
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Fixed issue where stationary stations like K5SGD-D were showing multiple
positions due to minor GPS accuracy variations. The previous 15-meter
threshold was too strict for typical GPS drift which can vary 5-30 meters.
- Increased default threshold in GeoUtils from 10m to 50m
- Updated packet processor to use 50m threshold for movement detection
- Verified with K5SGD-D showing ~24m variations that are now ignored
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add nil check before canceling timer in PacketConsumer to prevent errors
when timer reference doesn't exist (fixes test failures)
- Simplify packet counter reset migration by removing unsupported event
triggers, keeping only manual reset function
- All tests now passing
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Fixed "ParseError does not implement the Access behaviour" error that
occurred during packet batch insert when packets contained ParseError
structs in their data_extended field.
Added special handling for Aprs.Types.ParseError structs in the
struct_to_map function to convert them to simple maps with their
error_code and message fields, preventing the Access behavior error
when the struct is accessed during processing.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
The memory monitoring was incorrectly using total VM memory instead of
per-process memory, causing false positives whenever any process in the
system allocated memory.
Changes:
- Changed memory monitoring to per-process instead of total VM memory
- Reduced thresholds to reasonable per-process values (50MB diff, 100MB total)
- Increased minor GC trigger from 100 to 1000 packets to prevent GC after
every normal batch
This prevents detecting memory changes from other processes and eliminates
the constant GC warnings while still maintaining proper memory management.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Reduced batch processing parameters to prevent excessive GC:
- Reduced batch_size from 500 to 100 packets
- Adjusted batch_timeout from 2000ms to 1000ms for faster processing
- Reduced max_demand from 750 to 300 to match smaller batches
- Updated hardcoded fallback from 500 to 100 in packet_consumer.ex
- Changed minor GC trigger from 500 to 100 packets to match batch size
These smaller batch sizes were proven to work well in production and
should significantly reduce the frequency of garbage collection while
maintaining good performance.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Increased memory thresholds from 10MB/20MB to 500MB/1GB since the server
has 15GB available memory. This prevents constant GC warnings while
maintaining memory safety.
Configured VM garbage collection settings in vm.args.eex:
- Enabled concurrent ports/sockets limit (+Q 65536)
- Set heap sizes and GC parameters optimized for better performance
- Added scheduler and binary heap settings
These changes will reduce unnecessary GC overhead and improve overall
application performance.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
The PacketPipelineSetup was attempting to subscribe to a single
PacketConsumer process that doesn't exist. The application uses
PacketConsumerPool which automatically handles subscriptions for
multiple consumer processes, making PacketPipelineSetup redundant
and causing startup failures.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Create PacketConsumerPool to run 3 parallel consumers (configurable)
- Increase max_demand from 250 to 750 total (250 per consumer)
- Fix timer management to properly cancel/restart on batch processing
- Configure proper GenStage subscriptions with backpressure control
- Allow unnamed consumers for pool usage
This 3x increase in processing capacity prevents the "Packet buffer full"
warnings by ensuring consumers can keep up with incoming packet rate.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
This commit introduces several major performance improvements and adds
robust error handling for malformed HTTP requests:
Performance Optimizations:
- Add database migration with BRIN indexes for time-series data and partial
indexes for recent queries, significantly improving query performance
- Create global StreamingPacketsPubSub system for real-time packet distribution
with geographic bounds filtering using ETS for fast lookups
- Refactor PacketConsumer to use Stream module for memory-efficient processing,
preventing memory accumulation during batch operations
- Implement dedicated BroadcastTaskSupervisor pool for async broadcast operations,
preventing GenServer blocking on I/O
- Increase database connection pool from 25 to 45 (production) and 15 to 30 (dev)
for better concurrency support
Error Handling:
- Add SentryFilter to prevent Bandit.HTTPError from missing Host headers from
cluttering Sentry (common with bots/scanners)
- Configure custom 400 Bad Request error handling
- Filter out common bot/scanner paths from error reporting
All changes include comprehensive test coverage following TDD practices.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Rename Gleam module from encoding_simple to encoding
- Move Gleam files from src/aprs/ to src/aprsme/ to match project namespace
- Create custom Mix task for Gleam compilation (lib/mix/tasks/gleam_compile.ex)
- Update EncodingUtils to wrap Gleam implementation instead of pure Elixir
- Add Gleam dependencies to mix.exs and configure build paths
- Update Mix aliases to include Gleam compilation in test and compile tasks
- Add Gleam support to GitHub Actions CI workflow with caching
- Add GLEAM_INTEGRATION.md documentation
- All 357 tests passing with Gleam integration
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Replace bracket notation (packet_data[key]) with Map.get/2 to avoid
UndefinedFunctionError when processing ParseError structs that don't
implement the Access behavior.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Restored original PacketConsumer implementation from before optimizations
- Reverted to fixed batch size of 50 packets (original value)
- Removed SystemMonitor and InsertOptimizer from application startup
- Restored original config with buffer size 1000 and batch timeout 1000ms
- Removed all dynamic batch sizing and optimization logic
The recent performance optimizations were causing packet buffer overflows
because the dynamic batch sizing was actually slowing down processing.
Reverting to the simpler, working implementation that processes packets
consistently without buffer overflows.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Increased packet buffer size from 1,000 to 10,000 to handle traffic spikes
- Removed InsertOptimizer dynamic batch sizing that was slowing processing
- Reverted to fixed batch size of 200 for consistent performance
- Simplified insert options to use static configuration
- Added telemetry metrics for buffer overflow and utilization monitoring
- Fixed unused variable warning in packet_consumer.ex
The root cause was the recent performance optimizations that introduced
dynamic batch sizing, which actually degraded performance and caused
the producer buffer to overflow when the consumer couldn't keep up.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Fixed InsertOptimizer returning map instead of keyword list for Ecto
- Increased batch size ranges to handle high load (100-800 packets)
- Improved consumer responsiveness:
- Reduced batch timeout from 1000ms to 500ms
- Process batches at 80% capacity for better throughput
- Reduced adjustment interval from 10s to 5s
- Added comprehensive logging for debugging buffer status
- Created integration tests for packet pipeline
- Coordinated batch sizing between SystemMonitor and InsertOptimizer
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Renamed predicate functions to follow Elixir conventions:
- is_finite? → finite?
- is_finite_number? → finite_number?
- is_finite_float? → finite_float?
- Fixed all function returns to ensure proper nil handling
- Applied mix format to all files
- All tests pass with no compilation warnings
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Security improvements:
- Added sanitize_numeric_string to remove dangerous characters
- Limited input string length to prevent DoS attacks (20 chars for numbers)
- Added is_finite? checks to prevent infinity/NaN values
- Validate coordinate ranges (lat: -90 to 90, lng: -180 to 180)
- Added safe_parse_coordinate with proper validation
- Updated to_float in EncodingUtils with security validations
- Protected against integer overflow in coordinate conversions
- Added sanitize_path_string for APRS path validation
All coordinate inputs from users are now:
1. Sanitized to remove injection characters
2. Length-limited to prevent resource exhaustion
3. Validated for finite values (no infinity/NaN)
4. Checked against valid geographic ranges
5. Given safe fallback defaults
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
