piker/docs/macos/compatibility-fixes.md

505 lines
15 KiB
Markdown

# macOS Compatibility Fixes for Piker/Tractor
This guide documents macOS-specific issues encountered when running `piker` on macOS and their solutions. These fixes address platform differences between Linux and macOS in areas like socket credentials, shared memory naming, and async runtime coordination.
## Table of Contents
1. [Socket Credential Passing](#1-socket-credential-passing)
2. [Shared Memory Name Length Limits](#2-shared-memory-name-length-limits)
3. [Shared Memory Cleanup Race Conditions](#3-shared-memory-cleanup-race-conditions)
4. [Async Runtime (Trio/AsyncIO) Coordination](#4-async-runtime-trioasyncio-coordination)
---
## 1. Socket Credential Passing
### Problem
On Linux, `tractor` uses `SO_PASSCRED` and `SO_PEERCRED` socket options for Unix domain socket credential passing. macOS doesn't support these constants, causing `AttributeError` when importing.
```python
# Linux code that fails on macOS
from socket import SO_PASSCRED, SO_PEERCRED # AttributeError on macOS
```
### Error Message
```
AttributeError: module 'socket' has no attribute 'SO_PASSCRED'
```
### Root Cause
- **Linux**: Uses `SO_PASSCRED` (to enable credential passing) and `SO_PEERCRED` (to retrieve peer credentials)
- **macOS**: Uses `LOCAL_PEERCRED` (value `0x0001`) instead, and doesn't require enabling credential passing
### Solution
Make the socket credential imports platform-conditional:
**File**: `tractor/ipc/_uds.py` (or equivalent in `piker` if duplicated)
```python
import sys
from socket import (
socket,
AF_UNIX,
SOCK_STREAM,
)
# Platform-specific credential passing constants
if sys.platform == 'linux':
from socket import SO_PASSCRED, SO_PEERCRED
elif sys.platform == 'darwin': # macOS
# macOS uses LOCAL_PEERCRED instead of SO_PEERCRED
# and doesn't need SO_PASSCRED
LOCAL_PEERCRED = 0x0001
SO_PEERCRED = LOCAL_PEERCRED # Alias for compatibility
SO_PASSCRED = None # Not needed on macOS
else:
# Other platforms - may need additional handling
SO_PASSCRED = None
SO_PEERCRED = None
# When creating a socket
if SO_PASSCRED is not None:
sock.setsockopt(SOL_SOCKET, SO_PASSCRED, 1)
# When getting peer credentials
if SO_PEERCRED is not None:
creds = sock.getsockopt(SOL_SOCKET, SO_PEERCRED, struct.calcsize('3i'))
```
### Implementation Notes
- The `LOCAL_PEERCRED` value `0x0001` is specific to macOS (from `<sys/un.h>`)
- macOS doesn't require explicitly enabling credential passing like Linux does
- Consider using `ctypes` or `cffi` for a more robust solution if available
---
## 2. Shared Memory Name Length Limits
### Problem
macOS limits POSIX shared memory names to **31 characters** (defined as `PSHMNAMLEN` in `<sys/posix_shm_internal.h>`). Piker generates long descriptive names that exceed this limit, causing `OSError`.
```python
# Long name that works on Linux but fails on macOS
shm_name = "piker_quoter_tsla.nasdaq.ib_hist_1m" # 39 chars - too long!
```
### Error Message
```
OSError: [Errno 63] File name too long: '/piker_quoter_tsla.nasdaq.ib_hist_1m'
```
### Root Cause
- **Linux**: Supports shared memory names up to 255 characters
- **macOS**: Limits to 31 characters (including leading `/`)
### Solution
Implement automatic name shortening for macOS while preserving the original key for lookups:
**File**: `piker/data/_sharedmem.py`
```python
import hashlib
import sys
def _shorten_key_for_macos(key: str) -> str:
'''
macOS has a 31 character limit for POSIX shared memory names.
Hash long keys to fit within this limit while maintaining uniqueness.
'''
# macOS shm_open() has a 31 char limit (PSHMNAMLEN)
# Use format: /p_<hash16> where hash is first 16 hex chars of sha256
# This gives us: / + p_ + 16 hex chars = 19 chars, well under limit
# We keep the 'p' prefix to indicate it's from piker
if len(key) <= 31:
return key
# Create a hash of the full key
key_hash = hashlib.sha256(key.encode()).hexdigest()[:16]
short_key = f'p_{key_hash}'
return short_key
class _Token(Struct, frozen=True):
'''
Internal representation of a shared memory "token"
which can be used to key a system wide post shm entry.
'''
shm_name: str # actual OS-level name (may be shortened on macOS)
shm_first_index_name: str
shm_last_index_name: str
dtype_descr: tuple
size: int # in struct-array index / row terms
key: str | None = None # original descriptive key (for lookup)
def __eq__(self, other) -> bool:
'''
Compare tokens based on shm names and dtype, ignoring the key field.
The key field is only used for lookups, not for token identity.
'''
if not isinstance(other, _Token):
return False
return (
self.shm_name == other.shm_name
and self.shm_first_index_name == other.shm_first_index_name
and self.shm_last_index_name == other.shm_last_index_name
and self.dtype_descr == other.dtype_descr
and self.size == other.size
)
def __hash__(self) -> int:
'''Hash based on the same fields used in __eq__'''
return hash((
self.shm_name,
self.shm_first_index_name,
self.shm_last_index_name,
self.dtype_descr,
self.size,
))
def _make_token(
key: str,
size: int,
dtype: np.dtype | None = None,
) -> _Token:
'''
Create a serializable token that uniquely identifies a shared memory segment.
'''
if dtype is None:
dtype = def_iohlcv_fields
# On macOS, shorten long keys to fit the 31-char limit
if sys.platform == 'darwin':
shm_name = _shorten_key_for_macos(key)
shm_first = _shorten_key_for_macos(key + "_first")
shm_last = _shorten_key_for_macos(key + "_last")
else:
shm_name = key
shm_first = key + "_first"
shm_last = key + "_last"
return _Token(
shm_name=shm_name,
shm_first_index_name=shm_first,
shm_last_index_name=shm_last,
dtype_descr=tuple(np.dtype(dtype).descr),
size=size,
key=key, # Store original key for lookup
)
```
### Key Design Decisions
1. **Hash-based shortening**: Uses SHA256 to ensure uniqueness and avoid collisions
2. **Preserve original key**: Store the original descriptive key in the `_Token` for debugging and lookups
3. **Custom equality**: The `__eq__` and `__hash__` methods ignore the `key` field to ensure tokens are compared by their actual shm properties
4. **Platform detection**: Only applies shortening on macOS (`sys.platform == 'darwin'`)
### Edge Cases to Consider
- Token serialization across processes (the `key` field must survive IPC)
- Token lookup in dictionaries and caches
- Debugging output (use `key` field for human-readable names)
---
## 3. Shared Memory Cleanup Race Conditions
### Problem
During teardown, shared memory segments may be unlinked by one process while another is still trying to clean them up, causing `FileNotFoundError` to crash the application.
### Error Message
```
FileNotFoundError: [Errno 2] No such file or directory: '/p_74c86c7228dd773b'
```
### Root Cause
In multi-process architectures like `tractor`, multiple processes may attempt to clean up shared resources simultaneously. Race conditions during shutdown can cause:
1. Process A unlinks the shared memory
2. Process B tries to unlink the same memory → `FileNotFoundError`
3. Uncaught exception crashes Process B
### Solution
Add defensive error handling to catch and log cleanup races:
**File**: `piker/data/_sharedmem.py`
```python
class ShmArray:
# ... existing code ...
def destroy(self) -> None:
'''
Destroy the shared memory segment and cleanup OS resources.
'''
if _USE_POSIX:
# We manually unlink to bypass all the "resource tracker"
# nonsense meant for non-SC systems.
shm = self._shm
name = shm.name
try:
shm_unlink(name)
except FileNotFoundError:
# Might be a teardown race where another process
# already unlinked it - this is fine, just log it
log.warning(f'Shm for {name} already unlinked?')
# Also cleanup the index counters
if hasattr(self, '_first'):
try:
self._first.destroy()
except FileNotFoundError:
log.warning(f'First index shm already unlinked?')
if hasattr(self, '_last'):
try:
self._last.destroy()
except FileNotFoundError:
log.warning(f'Last index shm already unlinked?')
class SharedInt:
# ... existing code ...
def destroy(self) -> None:
if _USE_POSIX:
# We manually unlink to bypass all the "resource tracker"
# nonsense meant for non-SC systems.
name = self._shm.name
try:
shm_unlink(name)
except FileNotFoundError:
# might be a teardown race here?
log.warning(f'Shm for {name} already unlinked?')
```
### Implementation Notes
- This fix is platform-agnostic but particularly important on macOS where the shortened names make debugging harder
- The warnings help identify cleanup races during development
- Consider adding metrics/counters if cleanup races become frequent
---
## 4. Async Runtime (Trio/AsyncIO) Coordination
### Problem
The `TrioTaskExited` error occurs when trio tasks are cancelled while asyncio tasks are still running, indicating improper coordination between the two async runtimes.
### Error Message
```
tractor._exceptions.TrioTaskExited: but the child `asyncio` task is still running?
>>
|_<Task pending name='Task-2' coro=<wait_on_coro_final_result()> ...>
```
### Root Cause
`tractor` uses "guest mode" to run trio as a guest in asyncio's event loop (or vice versa). The error occurs when:
1. A trio task is cancelled (e.g., user closes the UI)
2. The cancellation propagates to cleanup handlers
3. Cleanup tries to exit while asyncio tasks are still running
4. The `translate_aio_errors` context manager detects this inconsistent state
### Current State
This issue is **partially resolved** by the other fixes (socket credentials and shared memory), which eliminate the underlying errors that trigger premature cancellation. However, it may still occur in edge cases.
### Potential Solutions
#### Option 1: Improve Cancellation Propagation (Tractor-level)
**File**: `tractor/to_asyncio.py`
```python
async def translate_aio_errors(
chan,
wait_on_aio_task: bool = False,
suppress_graceful_exits: bool = False,
):
'''
Context manager to translate asyncio errors to trio equivalents.
'''
try:
yield
except trio.Cancelled:
# When trio is cancelled, ensure asyncio tasks are also cancelled
if wait_on_aio_task:
# Give asyncio tasks a chance to cleanup
await trio.lowlevel.checkpoint()
# Check if asyncio task is still running
if aio_task and not aio_task.done():
# Cancel it gracefully
aio_task.cancel()
# Wait briefly for cancellation
with trio.move_on_after(0.5): # 500ms timeout
await wait_for_aio_task_completion(aio_task)
raise # Re-raise the cancellation
```
#### Option 2: Proper Shutdown Sequence (Application-level)
**File**: `piker/brokers/ib/api.py` (or similar broker modules)
```python
async def load_clients_for_trio(
client: Client,
...
) -> None:
'''
Load asyncio client and keep it running for trio.
'''
try:
# Setup client
await client.connect()
# Keep alive - but make it cancellable
await trio.sleep_forever()
except trio.Cancelled:
# Explicit cleanup before propagating cancellation
log.info("Shutting down asyncio client gracefully")
# Disconnect client
if client.isConnected():
await client.disconnect()
# Small delay to let asyncio cleanup
await trio.sleep(0.1)
raise # Now safe to propagate
```
#### Option 3: Detection and Warning (Current Approach)
The current code detects the issue and raises a clear error. This is acceptable if:
1. The error is rare (only during abnormal shutdown)
2. It doesn't cause data loss
3. Logs provide enough info for debugging
### Recommended Approach
For **piker**: Implement Option 2 (proper shutdown sequence) in broker modules where asyncio is used.
For **tractor**: Consider Option 1 (improved cancellation propagation) as a library-level enhancement.
### Testing
Test the fix by:
```python
# Test graceful shutdown
async def test_asyncio_trio_shutdown():
async with open_channel_from(...) as (first, chan):
# Do some work
await chan.send(msg)
# Trigger cancellation
raise KeyboardInterrupt
# Should cleanup without TrioTaskExited error
```
---
## Summary of Changes
### Files Modified in Piker
1. **`piker/data/_sharedmem.py`**
- Added `_shorten_key_for_macos()` function
- Modified `_Token` class to store original `key`
- Modified `_make_token()` to use shortened names on macOS
- Added `FileNotFoundError` handling in `destroy()` methods
2. **`piker/ui/_display.py`**
- Removed assertion that checked for 'hist' in shm name (incompatible with shortened names)
### Files to Modify in Tractor (Recommended)
1. **`tractor/ipc/_uds.py`**
- Make socket credential imports platform-conditional
- Handle macOS-specific `LOCAL_PEERCRED`
2. **`tractor/to_asyncio.py`** (Optional)
- Improve cancellation propagation between trio and asyncio
- Add graceful shutdown timeout for asyncio tasks
### Platform Detection Pattern
Use this pattern consistently:
```python
import sys
if sys.platform == 'darwin': # macOS
# macOS-specific code
pass
elif sys.platform == 'linux': # Linux
# Linux-specific code
pass
else:
# Other platforms / fallback
pass
```
### Testing Checklist
- [ ] Test on macOS (Darwin)
- [ ] Test on Linux
- [ ] Test shared memory with names > 31 chars
- [ ] Test multi-process cleanup race conditions
- [ ] Test graceful shutdown (Ctrl+C)
- [ ] Test abnormal shutdown (kill signal)
- [ ] Verify no memory leaks (check `/dev/shm` on Linux, `ipcs -m` on macOS)
---
## Additional Resources
- **macOS System Headers**:
- `/usr/include/sys/un.h` - Unix domain socket constants
- `/usr/include/sys/posix_shm_internal.h` - Shared memory limits
- **Python Documentation**:
- [`socket` module](https://docs.python.org/3/library/socket.html)
- [`multiprocessing.shared_memory`](https://docs.python.org/3/library/multiprocessing.shared_memory.html)
- **Trio/AsyncIO**:
- [Trio Guest Mode](https://trio.readthedocs.io/en/stable/reference-lowlevel.html#using-guest-mode-to-run-trio-on-top-of-other-event-loops)
- [Tractor Documentation](https://github.com/goodboy/tractor)
---
## Contributing
When implementing these fixes in your own project:
1. **Test thoroughly** on both macOS and Linux
2. **Add platform guards** to prevent cross-platform breakage
3. **Document platform-specific behavior** in code comments
4. **Consider CI/CD** testing on multiple platforms
5. **Handle edge cases** gracefully with proper logging
If you find additional macOS-specific issues, please contribute to this guide!