Skip to main content
As your infrastructure grows, optimizing pyinfra’s performance becomes crucial. This guide covers techniques to speed up deployments, reduce resource usage, and scale to large numbers of hosts.

​
Understanding pyinfra Performance

pyinfra’s performance is affected by:
  • Number of hosts: Operations execute in parallel across hosts
  • Number of operations: Each operation involves fact gathering and command execution
  • Network latency: SSH connections and command execution time
  • Fact gathering: Frequent fact queries can slow deployments
  • Operation complexity: Complex operations with many conditionals

​
Parallel Execution

pyinfra uses gevent for concurrent execution across hosts.

​
Controlling Parallelism

Adjust the number of parallel operations: 
# Default: 10 parallel hosts
pyinfra inventory.py deploy.py

# Increase to 50 parallel hosts
pyinfra --parallel 50 inventory.py deploy.py

# Reduce to 5 for resource-constrained systems
pyinfra --parallel 5 inventory.py deploy.py
In Python API: 
from pyinfra import Config, State
from pyinfra.api import Inventory

config = Config(
    PARALLEL=50,  # Execute on 50 hosts simultaneously
)

state = State(config=config, inventory=inventory)

​
Optimal Parallel Settings

Rules of thumb:
  • Small clusters (< 10 hosts): Use default (10)
  • Medium clusters (10-100 hosts): Set to 20-50
  • Large clusters (> 100 hosts): Set to 50-100
  • Very large clusters (> 1000 hosts): Consider batching (see below)

​
Fact Caching

Facts are cached per deployment, but repeated queries in operations can still be slow.

​
Avoid Repeated Fact Queries

Bad - queries fact multiple times: 
from pyinfra import host
from pyinfra.api import operation
from pyinfra.facts.server import Hostname

@operation()
def bad_example():
    # Queries hostname 3 times!
    if host.get_fact(Hostname) == "web1":
        yield f"echo {host.get_fact(Hostname)}"
    yield f"hostname {host.get_fact(Hostname)}"
Good - queries fact once: 
@operation()
def good_example():
    # Query once, reuse result
    hostname = host.get_fact(Hostname)
    
    if hostname == "web1":
        yield f"echo {hostname}"
    yield f"hostname {hostname}"

​
Preload Facts

For operations that always need certain facts, query them upfront: 
from pyinfra import host, State
from pyinfra.api import operation
from pyinfra.facts.files import Directory
from pyinfra.facts.server import Hostname, Os

# Preload facts for all hosts
def preload_common_facts(state: State):
    """Load commonly used facts upfront."""
    from pyinfra.api.facts import get_facts
    
    # Load facts in parallel across all hosts
    get_facts(
        state,
        Hostname,
    )
    get_facts(
        state,
        Os,
    )

# In deploy script
preload_common_facts(state)

# Now operations can use cached facts
@operation()
def optimized_operation():
    hostname = host.get_fact(Hostname)  # Uses cache
    os_name = host.get_fact(Os)          # Uses cache
    # ...

​
Batch Operations

For very large deployments, batch hosts into groups: 
from pyinfra import State
from pyinfra.api import Inventory

inventory = Inventory.load_inventory("large_inventory.py")

# Split into batches of 100 hosts
batch_size = 100
for i in range(0, len(inventory.hosts), batch_size):
    batch_hosts = list(inventory.hosts.values())[i:i+batch_size]
    
    # Create inventory with just this batch
    batch_inventory = Inventory(
        (batch_hosts, {}),
        override_data={}
    )
    
    state = State(inventory=batch_inventory, config=config)
    # Run operations on this batch
    # ...

​
Connection Reuse

SSH connections are expensive. Reuse them where possible.

​
SSH ControlMaster

Enable SSH connection multiplexing: 
# In ~/.ssh/config
Host *
    ControlMaster auto
    ControlPath ~/.ssh/control-%r@%h:%p
    ControlPersist 10m
This reuses a single SSH connection for multiple commands, drastically reducing connection overhead.

​
Keep Connections Alive

Set SSH keep-alive to prevent connection timeouts: 
from pyinfra import Config

config = Config(
    CONNECT_TIMEOUT=30,
    # SSH keep-alive (sent every 60 seconds)
    SSH_PARAMIKO_CONNECT_KWARGS={
        'timeout': 30,
        'banner_timeout': 30,
    },
)

​
Optimize Operations

​
Minimize Commands

Combine multiple commands into one: Bad - three separate commands: 
@operation()
def bad_example():
    yield "mkdir -p /opt/app"
    yield "chown app:app /opt/app"
    yield "chmod 755 /opt/app"
Good - one command: 
@operation()
def good_example():
    yield "mkdir -p /opt/app && chown app:app /opt/app && chmod 755 /opt/app"

​
Use Idempotent Checks

Skip operations that don’t need to run: 
from pyinfra import host
from pyinfra.api import operation
from pyinfra.facts.files import Directory

@operation()
def ensure_directory(path: str, user: str, mode: str):
    """Create directory only if it doesn't exist."""
    dir_info = host.get_fact(Directory, path=path)
    
    if dir_info is None:
        # Directory doesn't exist, create it
        yield f"mkdir -p {path}"
        yield f"chown {user}:{user} {path}"
        yield f"chmod {mode} {path}"
    elif dir_info["user"] != user or dir_info["mode"] != int(mode, 8):
        # Directory exists but wrong permissions
        yield f"chown {user}:{user} {path}"
        yield f"chmod {mode} {path}"
    else:
        # Already correct, do nothing
        host.noop(f"directory {path} already configured correctly")

​
Reduce Logging Output

Logging can slow down deployments with many operations.

​
Adjust Log Level

# Reduce verbosity
pyinfra --quiet inventory.py deploy.py

# Show only errors
pyinfra --log-level ERROR inventory.py deploy.py
In code: 
from pyinfra import Config

config = Config(
    QUIET=True,           # Minimal output
    LOG_LEVEL='WARNING',  # Only warnings and errors
)

​
Disable Fact Output

config = Config(
    PRINT_FACT_INFO=False,    # Don't print "Loaded fact..." messages
    PRINT_FACT_INPUT=False,   # Don't print fact commands
    PRINT_FACT_OUTPUT=False,  # Don't print fact output
)

​
File Transfer Optimization

​
Use rsync for Large Files

pyinfra supports rsync for efficient file transfers: 
from pyinfra.operations import files

files.rsync(
    name="Sync large directory",
    src="/local/large/dir/",
    dest="/remote/large/dir/",
    # Use compression for slow networks
    flags=["-az"],
)

​
Compress Files Before Transfer

from pyinfra.operations import files

# Compress locally
files.download(
    name="Download compressed archive",
    src="https://example.com/large-file.tar.gz",
    dest="/tmp/large-file.tar.gz",
)

# Extract remotely
server.shell(
    name="Extract archive",
    commands=["tar -xzf /tmp/large-file.tar.gz -C /opt/app"],
)

​
Memory Optimization

For deployments with many hosts, memory usage can be significant.

​
Limit Stored Output

By default, all command output is stored in memory: 
from pyinfra import Config

config = Config(
    # Don't store operation output
    SAVE_OUTPUT=False,
)

​
Clean Up After Operations

Delete temporary files during deployment: 
from pyinfra.operations import files, server

# Upload and process
files.put(
    name="Upload large file",
    src="large-file.dat",
    dest="/tmp/large-file.dat",
)

server.shell(
    name="Process file",
    commands=["process /tmp/large-file.dat > /opt/output.dat"],
)

# Clean up immediately
files.file(
    name="Remove temporary file",
    path="/tmp/large-file.dat",
    present=False,
)

​
Profiling Deployments

​
Time Individual Operations

Add timing to your deploy script: 
import time
from pyinfra import logger

def timed_operation(name: str, operation_func, *args, **kwargs):
    """Execute operation and log execution time."""
    start = time.time()
    result = operation_func(*args, **kwargs)
    elapsed = time.time() - start
    logger.info(f"{name} took {elapsed:.2f}s")
    return result

# Use in deploy script
timed_operation(
    "Install packages",
    apt.packages,
    packages=["nginx", "postgresql"],
    update=True,
)

​
Deployment Summary

After deployment, review the summary: 
pyinfra inventory.py deploy.py

# Output includes timing:
# --> Complete! Took 45.2s
# --> Operations: 127 (45 changes, 82 no change)
# --> Hosts: 10 (10 success, 0 failed)

​
Caching Strategy

​
Cache Expensive Operations

For operations that rarely change, cache their results: 
from functools import lru_cache
from pyinfra import host
from pyinfra.facts.server import Which

@lru_cache(maxsize=None)
def get_package_manager():
    """Detect package manager (cached)."""
    if host.get_fact(Which, command="apt-get"):
        return "apt"
    elif host.get_fact(Which, command="yum"):
        return "yum"
    elif host.get_fact(Which, command="pacman"):
        return "pacman"
    return None

# First call queries facts
pm = get_package_manager()  

# Subsequent calls use cache
pm = get_package_manager()

​
Network Optimization

​
Reduce Round Trips

Minimize commands that require remote state checks: 
# Bad - checks state for each file
for config_file in config_files:
    files.template(
        src=f"templates/{config_file}",
        dest=f"/etc/app/{config_file}",
    )

# Good - upload all at once
files.rsync(
    src="templates/",
    dest="/etc/app/",
)

​
Use Local Execution

For operations that don’t need remote execution: 
from pyinfra.operations import local

# Run locally on controller
local.shell(
    name="Generate config locally",
    commands=["./generate-config.sh"],
)

# Then upload
files.put(
    src="generated-config.yml",
    dest="/etc/app/config.yml",
)

​
Database and Service Operations

​
Batch Database Operations

from pyinfra.operations import postgresql

# Bad - separate transaction for each
for user in users:
    postgresql.user(
        user=user["name"],
        password=user["password"],
    )

# Good - single SQL script
sql_script = "\n".join([
    f"CREATE USER {u['name']} WITH PASSWORD '{u['password']}';" 
    for u in users
])

postgresql.sql(
    name="Create all users",
    sql=sql_script,
)

​
Best Practices Summary

  1. Increase parallelism for large deployments (—parallel flag)
  2. Cache facts by querying once and reusing results
  3. Batch operations for very large host counts
  4. Enable SSH ControlMaster for connection reuse
  5. Combine commands to reduce round trips
  6. Use idempotent checks to skip unnecessary work
  7. Reduce logging in production deployments
  8. Use rsync for large file transfers
  9. Clean up temporary files to save memory
  10. Profile deployments to identify bottlenecks

​
Benchmarking Example

Compare before and after optimization: 
# Before optimization
# Time: 180s for 50 hosts
# Operations: 200

# After optimization:
# - Increased --parallel to 30
# - Enabled SSH ControlMaster
# - Cached common facts
# - Combined 20 commands into 5
# - Reduced logging

# Time: 45s for 50 hosts (4x faster)
# Operations: 180 (20 skipped as unnecessary)

​
Next Steps

Build docs developers (and LLMs) love

Get started for freeTalk to us