MQTT is an easily used networking protocol designed for IOT (internet of things) applications. It is well suited for controlling hardware devices and for reading sensors across a local network or the internet.
It is a means of communicating between multiple clients. A single server, also known as a broker, manages the network. Clients may include ESP8266, ESP32 and Pyboard D modules and other networked computers. Typical server hardware is a Raspberry Pi or other small Linux machine which may be left running 24/7. An effective PC server is mosquitto. Public brokers also exist.
MQTT Packets are passed between clients using a publish/subscribe model. They consist of a topic and a message string. Clients subscribe to a topic and will receive all packets published by any client under that topic.
The protocol supports three "quality of service" (qos) levels. Level 0 offers no guarantees. Level 1 ensures that a packet is communicated to the recipient but duplication can occur. Level 2 avoids duplication; it is not supported by the official driver or by this module. Duplicates can readily be handled at the application level.
V1.22.0 included a changed IDF version 5.0.4: on ESPx the package should be replaced with the latest version, otherwise recovery from an outage may not occur.
The official "robust" MQTT client has the following limitations.
It is unable reliably to resume operation after a temporary WiFi outage.
It uses blocking sockets which can cause execution to pause for arbitrary periods when accessing a slow broker. It can also block forever in the case of qos == 1 publications while it waits for a publication acknowledge which never arrives; this can occur on a WiFi network if an outage occurs at this point in the sequence.
This blocking behaviour implies limited compatibility with asynchronous applications since pending coroutines will not be scheduled for the duration.
Its support for qos == 1 is partial. It does not support retransmission in the event of a publication acknowledge being lost. This can occur on a WiFi network especially near the limit of range or in the presence of interference.
Its partial qos == 1 support and inability reliably to resume after a WiFi outage places a limit on the usable WiFi range. To achieve reliable operation a client must be well within range of the access point (AP).
As a synchronous solution it has no mechanism to support the "keepalive" mechanism of MQTT. This prevents the "last will" system from working. It also makes subscribe-only clients problematic: the broker has no means of "knowing" whether the client is still connected.
This module aims to address these issues, at the cost of significant code size. It has been tested on the following platforms.
The principal features of this driver are:
qos == 1 operation with retransmission.It has the drawback of increased code size which is an issue on the ESP8266. Run as frozen bytecode it uses about 50% of the RAM on the ESP8266. On ESP32 and Pyboard D it may be run as a standard Python module.
This module provides a "resilient" non-blocking MQTT driver. In this context "resilient" means capable of reliable operation in the presence of poor WiFi connectivity and dropouts. Clearly during a dropout or broker outage communication is impossible but when connectivity resumes the driver recovers transparently.
Near the limit of WiFi range communication delays may be incurred owing to retransmissions and reconnections but nonblocking behaviour and qos == 1 integrity are maintained.
It supports qos levels 0 and 1. In the case of packets with qos == 1 retransmissions will occur until the packet has successfully been transferred. If the WiFi fails (e.g. the device moves out out of range of the AP) the coroutine performing the publication will pause until connectivity resumes.
The driver requires the asyncio library and is intended for applications
that use it. It uses nonblocking sockets and does not block the scheduler. The
design is based on the official umqtt library but it has been substantially
modified for resilience and for asynchronous operation.
It is primarily intended for applications which open a link to the MQTT broker aiming to maintaining that link indefinitely. Applications which close and re-open the link (e.g. for power saving purposes) are subject to limitations detailed in Non standard applications.
Hardware support: Pyboard D, ESP8266, ESP32, ESP32-S3, ESP32-S2, Pico W and
Arduino Nano RP2040 Connect.
Firmware support: Official MicroPython firmware V1.19 or later.
Broker support: Mosquitto is preferred for its excellent MQTT compliance.
Protocol: The module supports a subset of MQTT revision 3.1.1.
Initial development was by Peter Hinch. Thanks are due to Kevin Köck for providing and testing a number of bugfixes and enhancements. Also to other contributors, some mentioned below.
Note that in firmware prior to 1.21 asyncio was named uasyncio.
24 Oct 2024 V0.8.2 Socket reads use pre-allocated buffer for performance.
18 Aug 2024 V0.8.1 Reconfigured as a Python package. Bugfix in V5 support.
9 Aug 2024 V0.8.0 Partial MQTTv5 support contributed by Bob Veringa.
15 Feb 2024 V0.7.2 Make compliant with firmware V1.22.0 and later.
12 Nov 2022 V0.7.0 Provide alternative callback-free Event interface.
2 Nov 2022 Rename config.py to mqtt_local.py, doc improvements.
8 Aug 2022 V0.6.6 Support unsubscribe (courtesy of Kevin Köck's fork).
11 July 2022 V0.6.5 Support RP2 Pico W
5 July 2022 V0.6.4 Implement enhancements from Bob Veringa. Fix bug where tasks
could fail to be stopped on a brief outage. Subscription callbacks now receive
bytearrays rather than bytes objects.
10 June 2022 Lowpower demo removed as it required an obsolete version of
asyncio. Improved handling of clean_init (issue #40).
21 May 2022 SSL/TLS ESP8266 support contributed by @SooOverpowered: see
tls8266.py.
22 Apr 2022 Support added for Arduino Nano RP2040 Connect. See note below.
2 Aug 2021 SSL/TLS on ESP32 has now been confirmed working.
Reference.
The module is too large to compile on the ESP8266 and should be precompiled or
preferably frozen as bytecode. On the reference board with mqtt_as frozen,
the demo script range_ex reports 27.4K of free RAM while running. The code
disables automatic sleep: this reduces reconnects at cost of increased power
consumption.
Notes on the Sonoff Basic R3 may be found here.
Firmware must now be official firmware as described above. The Loboris port has been abandoned by its author and is no longer supported.
The library has been tested successfully with the Pyboard D SF2W and SF6W. In testing it has clocked up eight weeks of continuous runtime and nearly 1M messages without failure or data loss.
NINA firmware must up to date otherwise MicroPython produces error messages.
See
this doc.
Reading RSSI seems to break the WiFi link so should be avoided - the
range_ex.py demo disables this on this platform.
The mqtt_as code should be V0.6.5 or later to avoid very slow recovery from
outages.
The MQTT 3.1 protocol supports extremely long messages. On a microcontroller message length is limited by available RAM. The actual limit will depend on the platform and user code but it is wise to design on the basis of a maximum of around 1KiB.
Avoid unrealistic expectations of performance: latency can be significant, especially when using a TLS connection to a broker located on the internet. With a non-encrypted connection to a local broker it is feasible to use one MicroPython client to control another. I haven't measured latency but I would guess at ~100ms.
Some platforms - notably ESP32 - are unhelpful when dealing with gross errors such as incorrect WiFi credentials. Initial connection will only fail after a one minute timeout. Other platforms enable an immediate bail-out.
The addition of MQTTv5 support does not affect existing applications which will run unchanged. It is expected that most microcontroller users will continue with MQTT V3.1.1. The use of MQTTv5 uses additinal RAM (~3KiB) and requires some knowledge of the protocol. See MQTTv5 Support for more details.
The library is configured as a Python package and is installed to an mqtt_as
directory. This is typically located on the Python path. Demo scripts are then
run with (e.g.):
>>> import mqtt_as.rangeThe user-configured mqtt_local.py (see below) is located on the Python path.
__init__.py The main module.mqtt_v5_properties.py Only required if using MQTTv5.mqtt_local.py Holds local configuration details such as WiFi credentials.
Place on Python path (usually / or /lib/).The first two of these demonstrate the event interface. Others use callbacks.
range.py For WiFi range testing. Good general demo.range_ex.py As above but also publishes RSSI and free RAM. See code
comments for limitations on Pico W and Arduino nano connect.clean.py Test/demo program using MQTT Clean Session.unclean.py Test/demo program with MQTT Clean Session False.main.py Example for auto-starting an application.tls.py Demo of SSL/TLS connection to a public broker. This runs on a
Pyboard D. Publishes every 20s and subscribes to same topic. Connection to
this public broker, though encrypted, is insecure because anyone can
subscribe.tls8266.py SSL/TLS connection for ESP8266. Shows how to use keys and
certificates. For obvious reasons it requires editing to run.
Test scripts for MQTTv5:
basic.py Demo of user properties under MQTTv5.
Bash scripts (may be run on PC to publish periodically):
pubtest Bash script illustrating publication with Mosquitto.pubtest_v5 Bash script illustrates various publication properties.ESP8266: please read Installation. On other
platforms the main module, demos 1 to 3 and the sample mqtt_local.py may be
installed from a connected PC with :
$ mpremote mip install github:peterhinch/micropython-mqttFor MQTTv5, demos may be added with
$ mpremote mip install github:peterhinch/micropython-mqtt/mqtt_as/v5An alternative is to use mip at the REPL with WiFi connected:
>>> import mip
>>> mip.install("github:peterhinch/micropython-mqtt")The Bash scripts pubtest and pubtest_v5 should be copied to the PC.
The MQTT client is configured using a dictionary. An instance named config
is defined in the MQTTClient class and
populated with common default values. The user can populate this in any manner.
The approach used in the test scripts is as follows. The main __init__.py
module instantiates config with typical defaults. Then mqtt_local.py adds
local settings common to all nodes, e.g. WiFi credentials and broker details.
Finally the application adds application specific settings like subscriptions.
In a typical project mqtt_local.py will be edited then deployed to all nodes.
The ESP8266 stores WiFi credentials internally: if the ESP8266 has connected to
the LAN prior to running there is no need explicitly to specify these. On other
platforms, or to have the capability of running on an ESP8266 which has not
previously connected, mqtt_local.py should be edited to provide them. This is
a sample cross-platform file:
from mqtt_as import config
config['server'] = '192.168.0.10' # Change to suit e.g. 'iot.eclipse.org'
# Required on Pyboard D and ESP32. On ESP8266 these may be omitted (see above).
config['ssid'] = 'my_WiFi_SSID'
config['wifi_pw'] = 'my_password'The module is too large to compile on the ESP8266. It must either be cross
compiled or (preferably) built as frozen bytecode: copy __init__.py to
esp8266/modules in the source tree, build and deploy. Copy mqtt_local.py to
the filesystem for ease of making changes.
On other platforms simply copy the Python source to the filesystem (items 1 and 2 above as a minimum).
If an application is to auto-run on power-up it can be necessary to add a short delay in main.py:
import time
time.sleep(5) # Could probably be shorter
import range # Your applicationThis is platform dependent and gives the hardware time to initialise.
The library offers two alternative ways to handle events such as the arrival of
a message. One uses traditional callbacks. The following uses Event instances
and an asynchronous iterator. If a PC client publishes a message with the topic
foo_topic the topic and message are printed. The code periodically publishes
an incrementing count under the topic result.
from mqtt_as import MQTTClient, config
import asyncio
# Local configuration
config['ssid'] = 'your_network_name' # Optional on ESP8266
config['wifi_pw'] = 'your_password'
config['server'] = '192.168.0.10' # Change to suit e.g. 'iot.eclipse.org'
async def messages(client): # Respond to incoming messages
# If MQTT V5is used this would read
# async for topic, msg, retained, properties in client.queue:
async for topic, msg, retained in client.queue:
print(topic.decode(), msg.decode(), retained)
async def up(client): # Respond to connectivity being (re)established
while True:
await client.up.wait() # Wait on an Event
client.up.clear()
await client.subscribe('foo_topic', 1) # renew subscriptions
async def main(client):
await client.connect()
for coroutine in (up, messages):
asyncio.create_task(coroutine(client))
n = 0
while True:
await asyncio.sleep(5)
print('publish', n)
# If WiFi is down the following will pause for the duration.
await client.publish('result', '{}'.format(n), qos = 1)
n += 1
config["queue_len"] = 1 # Use event interface with default queue size
MQTTClient.DEBUG = True # Optional: print diagnostic messages
client = MQTTClient(config)
try:
asyncio.run(main(client))
finally:
client.close() # Prevent LmacRxBlk:1 errorsThe code may be tested by running pubtest in one terminal and, in another,
mosquitto_sub -h 192.168.0.10 -t result (change the IP address to match your
broker).
The alternative callback-based interface may be run as follows:
from mqtt_as import MQTTClient, config
import asyncio
# Local configuration
config['ssid'] = 'your_network_name' # Optional on ESP8266
config['wifi_pw'] = 'your_password'
config['server'] = '192.168.0.10' # Change to suit e.g. 'iot.eclipse.org'
def callback(topic, msg, retained, properties=None): # MQTT V5 passes properties
print((topic.decode(), msg.decode(), retained))
async def conn_han(client):
await client.subscribe('foo_topic', 1)
async def main(client):
await client.connect()
n = 0
while True:
await asyncio.sleep(5)
print('publish', n)
# If WiFi is down the following will pause for the duration.
await client.publish('result', '{}'.format(n), qos = 1)
n += 1
config['subs_cb'] = callback
config['connect_coro'] = conn_han
MQTTClient.DEBUG = True # Optional: print diagnostic messages
client = MQTTClient(config)
try:
asyncio.run(main(client))
finally:
client.close() # Prevent LmacRxBlk:1 errorsAs above, testing is done by running pubtest in one terminal and, in another,
mosquitto_sub -h 192.168.0.10 -t result (change the IP address to match your
broker).
The module provides a single class: MQTTClient.
This takes a dictionary as argument. The default is mqtt_as.config which is
populated with default values listed below. A typical application imports this
and modifies selected entries as required. Entries are as follows (default
values shown in []):
These are required for platforms other than ESP8266 where they are optional. If the ESP8266 has previously connected to the required LAN the chip can reconnect automatically. If credentials are provided, an ESP8266 which has no stored values or which has stored values which don't match any available network will attempt to connect to the specified LAN.
'ssid' [None]
'wifi_pw' [None]
'client_id' [auto-generated unique ID] Must be a bytes instance.
'server' [None] Broker IP address (mandatory).
'port' [0] 0 signifies default port (1883 or 8883 for SSL).
'user' [''] MQTT credentials (if required).
'password' [''] If a password is provided a user must also exist.
'keepalive' [60] Period (secs) before broker regards client as having died.
'ping_interval' [0] Period (secs) between broker pings. 0 == use default.
'ssl' [False] If True use SSL.
'ssl_params' [{}] See below.
'response_time' [10] Time in which server is expected to respond (s). See note
below.
'clean_init' [True] Clean Session state on initial connection. (Ignored if
MQTT V5 is in use).
'clean' [True] Clean session state on reconnection. (Known as Clean Start in MQTT V5).
'max_repubs' [4] Maximum no. of republications before reconnection is
attempted.
'will' : [None] A list or tuple defining the last will (see below).
'queue_len' [0] If a value > 0 is passed the Event-based interface is
engaged. This replaces the callbacks defined below with a message queue and
Event instances. See section 3.5.
This interface is optional. It is retained for compatibility with existing code. In new designs please consider the event based interface which replaces callbacks with a more asyncio-friendly approach.
'subs_cb' [a null lambda function] Subscription callback. Runs when a message
is received whose topic matches a subscription. The callback must take three or
four args, topic, message, retained and properties=None. The first two
are bytes instances, retained is a bool, True if the message is a
retained message. properties is a dict (or None) if MQTT V5 is in use.
'wifi_coro' [a null coro] A coroutine. Defines a task to run when the network
state changes. The coro receives a single bool arg being the network state.
'connect_coro' [a null coro] A coroutine. Defines a task to run when a
connection to the broker has been established. This is typically used to
register and renew subscriptions. The coro receives a single argument, the
client instance.
See MQTTv5 Support
'mqttv5' [False]
'mqttv5_con_props' [None]
The response_time entry works as follows. If a read or write operation times
out, the connection is presumed dead and the reconnection process begins. If a
qos==1 publication is not acknowledged in this period, republication will
occur. May need extending for slow internet connections.
The will entry defines a publication which the broker will issue if it
determines that the connection has timed out. This is a tuple or list comprising
[topic (string), msg (string), retain (bool), qos (0 or 1)]. If the arg
is provided all elements are mandatory.
Clean sessions: If clean is set, messages from the server during an outage
will be lost regardless of their qos level.
If clean is False messages sent from the server with qos == 1 will be
received when connectivity resumes. This is standard MQTT behaviour (MQTT spec
section 3.1.2.4). If the outage is prolonged this can imply a substantial
backlog. On the ESP8266 this can cause buffer overflows in the Espressif WiFi
stack causing LmacRxBlk:1 errors to appear.
see this doc.
clean_init should normally be True. If False the system will attempt to
restore a prior session on the first connection. This may result in a large
backlog of qos==1 messages being received, for example if a client is taken
out of service for a long time. This can have the consequences described above.
See MQTT spec 3.1.2.4. This is described further below in
section 4.4.2 behaviour on power up.
Populating the ssl_params dictionary is something of a black art. Some sites
require certificates: see this post
for details on how to specify these. See Hive MQ for
details of connecting to a secure, free broker service. This may provide hints
for connecting to other TLS brokers.
Note re data types. Messages and topics may be strings provided that all
characters have ordinal values <= 127 (Unicode single byte characters).
Otherwise the string encode method should be used to convert them to bytes
objects.
Asynchronous.
Keyword only arg:
quick=False Setting quick=True saves power in some battery applications.
It does this on (re)connect by skipping the check on WiFi integrity. This check
is intended for mobile clients that may attempt to reconnect under conditions
of poor signal strength. In conditions of good signal strength this check may
be skipped.Connects to the specified broker. The application should call connect once on
startup. If this fails (due to WiFi or the broker being unavailable) an
OSError will be raised: see
Connect Error Codes. Subsequent
reconnections after outages are handled automatically.
Asynchronous.
If connectivity is OK the coro will complete immediately, else it will pause until the WiFi/broker are accessible. Section 4.2 describes qos == 1 operation.
Args:
topic A bytes or bytearray object. Or ASCII string as described above.msg A bytes or bytearray object.retain=False Boolean.qos=0 Integer.properties=None See MQTTv5 Support.Asynchronous.
Subscriptions should be created in the connect coroutine to ensure they are re-established after an outage.
The coro will pause until a SUBACK has been received from the broker, if
necessary reconnecting to a failed network.
Args:
topic A bytes or bytearray object. Or ASCII string as described above.qos=0 Integer.It is possible to subscribe to multiple topics but there can only be one subscription callback.
Asynchronous.
The coro will pause until an UNSUBACK has been received from the broker, if
necessary reconnecting to a failed network.
Arg:
topic A bytes or bytearray object. Or ASCII string as described above.If there is no subscription in place with the passed topic name the method will complete normally. This is in accordance with MQTT spec 3.10.4 Response.
Synchronous. No args.
Returns True if connectivity is OK otherwise it returns False and schedules
reconnection attempts.
Asynchronous. No args.
Sends a DISCONNECT packet to the broker, closes socket. Disconnection
suppresses the Will (MQTT spec. 3.1.2.5). This may be done prior to a power
down or deepsleep. For restrictions on the use of this method see
lightsleep and disconnect.
Synchronous. No args.
Shuts down the WiFi interface and closes the socket. Its main use is in
development to prevent ESP8266 LmacRxBlk:1 failures if an application raises
an exception or is terminated with ctrl-C (see
Example Usage.
Asynchronous. No args.
Unless data was received in the last second it issues an MQTT ping and waits
for a response. If it times out (response_time exceeded) with no response it
returns False otherwise it returns True.
Asynchronous.
Returns True if internet connectivity is available, else False. It first
checks current WiFi and broker connectivity. If present, it sends a DNS query
to '8.8.8.8' and checks for a valid response.
There is a single arg packet which is a bytes object being the DNS query. The
default object queries the Google DNS server.
Please note that this is merely a convenience method. It is not used by the client code and its use is entirely optional.
If the class variable DEBUG is true, debug messages are output via dprint.
This method can be redefined in a subclass, for example to log debug output to
a file. The method takes an arbitrary number of positional args as per print.
DEBUG If True causes diagnostic messages to be printed.REPUB_COUNT For debug purposes. Logs the total number of republications
with the same PID which have occurred since startup.VERSION A 3-tuple of ints (major, minor, micro) e.g. (0, 5, 0).This is invoked by setting config["queue_len"] = N where N > 0. In this
mode there are no callbacks. Incoming messages are queued and may be accessed
with an asynchronous iterator. The module reports connectivity changes by
setting bound .up and .down Event instances. The demos range.py and
range_ex.py use this interface. The following code fragments illustrate its use.
Reading messages:
async def messages(client):
# If MQTT V5 is in use
# async for topic, msg, retained, properties in client.queue:
async for topic, msg, retained in client.queue:
print(f'Topic: "{topic.decode()}" Message: "{msg.decode()}" Retained: {retained}')Handling connect events:
async def up(client): # (re)connection.
while True:
await client.up.wait()
client.up.clear()
print('We are connected to broker.')
await client.subscribe('foo_topic', 1) # Re-subscribe after outageOptional outage handler:
async def down(client):
while True:
await client.down.wait() # Pause until outage
client.down.clear()
print('WiFi or broker is down.')Initialisation with a default small queue:
config["queue_len"] = 1
client = MQTTClient(config)In applications where incoming messages arrive slowly and the clean flag is
False, messages will not accumulate and the queue length can be small. A
value of 1 provides a minimal queue. The queue comes into play when bursts of
messages arrive too quickly for the application to process them. This can occur
if multiple clients independently publish to the same topic: the broker may
forward them to subscribers at a high rate. Another case is when the clean
flag is False and a long wifi outage occurs: when the outage ends there may
be a large backlog of messages. Such cases may warrant a larger queue.
In the event of the queue overflowing, the oldest messages will be discarded.
This policy prioritises resilience over the qos==1 guarantee. The bound
variable client.queue.discards keeps a running total of lost messages. In
development this can help determine the optimum queue length.
It is possible (though seldom useful) to have multiple tasks waiting on messages. These must yield control after each message to allow the others to be scheduled. Messages will be distributed between waiting tasks in a round-robin fashion. Multiple instances of this task may be created:
async def messages(client):
async for topic, msg, retained in client.queue:
await asyncio.sleep(0) # Allow other instances to be scheduled
# handle messageIn applications RAM is at a premium, in testing the callback-based interface offers somewhat (~1.3KiB) lower consumption than the minimal queue case.
Application designers should consider whether V5 is appropriate. Relevant factors include:
False.
The intention here is to ensure that, where connectivity is unreliable, messages
are not lost during brief outages. In V3.1.1 this has the unfortunate
consequence that, where a prolonged outage occurs, there can be a large backlog.
V5 messages may be published with a "message expiry interval" property, offering
an option to limit the size of the backlog. This is further enhanced by a
"session expiry interval" set on connection.MQTTv5 is supported and can be configured by setting mqttv5 to True in the
the config dictionary. The default is False. V5 applications should not use
the clean_init config value - message backlogs should be controlled using
session and message expiry intervals. Properties on connect are supported, these
need to be passed in the configuration dictionary. See 3.6.2 for more
information on properties, and how to format them.
from mqtt_as import MQTTClient, config
config['mqttv5'] = True
# Optional: Set the properties for the connection
config['mqttv5_con_props'] = {
0x11: 3600, # Session Expiry Interval
}
# The rest of the configuration
client = MQTTClient(config)There are modifications to the API to support MQTTv5 features. The most
significant is the addition of the properties argument that is provided as
an additional argument to both the event and callback-based message handlers.
# For MQTT 3.1.1 support
def callback(topic, msg, retained):
print((topic, msg, retained))
# For MQTT 5 and 3.1.1 support
def callback(topic, msg, retained, properties=None):
print((topic, msg, retained, properties))Allowing properties as an optional argument allows you to switch between MQTT 3.1.1 and MQTT 5 support without changing the callback signature.
async def messages(client):
async for topic, msg, retained, properties in client.queue:
await asyncio.sleep(0) # Allow other instances to be scheduled
# handle messageThe properties argument is a dictionary that contains the properties of the
message. The properties are defined in the MQTTv5 specification. If you
include properties in published messages, while using MQTTv3.1.1, the properties
will be ignored.
Properties are a new and important feature of MQTTv5. They are used to provide additional information about the message, and allow for more advanced features such as message expiry, user properties, and response information.
Incoming properties are formatted as a dictionary using the property identifier as the key. The property identifier is an integer that is defined in the MQTTv5 specification, there are no constants defined in the module for these values. The property identifier is defined in the MQTTv5 specification.
Sending properties must be done in the right format. The MQTTv5 specification makes a distinction between binary and text properties. It is important to ensure that the properties are sent in the correct format. For reference, refer to section 2.2.2.2 of the MQTTv5 specification.
properties = {
0x26: {'value': 'test'}, # User Property (UTF-8 string pair)
0x09: b'correlation_data', # Correlation Data (binary)
0x08: 'response_topic', # Response Topic (UTF-8 string)
0x02: 60, # Message Expiry Interval (integer)
}
await client.publish('topic/test', 'message', False, 0, properties=properties)In the following tables of properties types are defined as Python variable
types; "string" is a utf8-encoded str. The V5 protocol provides for an
optional request/response exchange. This is described in the V5 specification
section 4.10
The following is a summary of properties relevant to client.publish(). |
Key | Value | Destination | Name | Meaning |
|---|---|---|---|---|---|
| 0x01 | byte | subscriber | payload format indicator | 0=binary 1=utf8 | |
| 0x02 | int | broker | Message Expiry Interval | Lifetime in seconds | |
| 0x03 | string | subscriber | Content Type | Application defined | |
| 0x08 | string | subscriber | Response Topic | Request/response | |
| 0x09 | bytes | subscriber | Correlation Data | Request/response | |
| 0x23 | int | broker | topic alias | ||
| 0x26 | string pair | subscriber | user property | Application defined |
Properties relevant to client.subscribe(): |
Key | Value | Name | Meaning |
|---|---|---|---|---|
| 0x0B | int | Subscription Identifier | See below | |
| 0x26 | string pair | user property | Application defined |
The subscription identifier enables a client application to pass its current state to the broker: responses to that subscription will include that state. See spec section 3.8.4
Properties relevant to client.connect(). Note that connection properties are
provided in the configuration dictionary (config[mqttv5_con_props]). All are
instructions to the broker. All except 0x11 and 0x22 are somewhat esoteric and
require study of the spec.
| Key | Value | Name | Meaning |
|---|---|---|---|
| 0x11 | int | Session expiry interval (secs) | See below |
| 0x17 | byte 0/1 | Request Problem Information | Spec 3.1.2.11.7 |
| 0x19 | byte 0/1 | Request Response Information | Spec 3.1.2.11.6 |
| 0x21 | int | Receive Maximum | Spec section 4.9 etc |
| 0x22 | int | Topic Alias Maximum | Max alias value |
| 0x26 | string pair | user property | Application defined |
| 0x27 | int | Maximum Packet Size | Spec 3.2.2.3.6 |
The Session Expiry Interval defines how long the broker will retain session
state after a disconnection. Setting it to zero and setting clean=True is
equivalent to Clean Session status in MQTT3.1.1: publications occurring during
the outage will be missed. A long expiry interval enables such messages to be
received, at risk of a large backlog after a prolonged outage.
| Incoming message: | Key | Value | Source | Name | Meaning |
|---|---|---|---|---|---|
| 0x01 | byte | publisher | payload format indicator | 0=binary 1=utf8 | |
| 0x02 | int | publisher | Message Expiry Interval | Lifetime in seconds | |
| 0x03 | string | publisher | Content Type | Application defined | |
| 0x08 | string | publisher | Response Topic | Request/response | |
| 0x09 | bytes | publisher | Correlation Data | Request/response | |
| 0x0B | int | publisher | Subscription Identifier | ||
| 0x26 | string pair | publisher | user property | Application defined |
Other packets received from the broker may contain properties. Apart from
CONNACK the mosquitto broker seems to send these only under error
conditions. In any event the client prints these if MQTTClient.DEBUG is
True. Example packets are CONNACK, PUBACK, SUBACK, UNSUBACK and
DISCONNECT.
The idea behind the topic alias is to reduce outgoing message size. A
publication is made with the full topic name with the topic alias set to a
nonzero integer. Subsequent publications may pass a topic of "" with the
topic alias property set to that integer. It is essential that the broker
receives the message setting the alias as it can disconnect if an unknown alias
is received. A problem also arises if an outage occurs while publishing aliased
messages. It seems that the broker does not store aliases after an outage. It is
the responsibility of the application to re-establish any aliases on reconnect.
Please study
the spec
before using this feature.
In the interest of keeping the library lightweight and well tested, some features of MQTTv5 are not supported.
AUTH packet is not implemented and is not handled.CONNACK packet are exposed.CONNECT and PUBLISH are not
returned to the user. For more information, see this
commentNOTE: Most of these features could be implemented with some effort. These features were not implemented, to keep the current implementation simple and reduce the scope of testing required.
If keepalive is defined in the constructor call, the broker will assume that
connectivity has been lost if no messages have been received in that period.
The module attempts to keep the connection open by issuing an MQTT ping up to
four times during the keepalive interval. (It pings if the last response from
the broker was over 1/4 of the keepalive period). More frequent pings may be
desirable to reduce latency of outage detection. This may be done using the
ping_interval configuration option. The point here is that while WiFi
failures are detected fast, upstream failure can only be detected by an absence
of communication from the broker. With a long ping interval, the broker could
be unreachable for a long time before the client detects it and initiates a
reconnection attempt.
If the broker times out it will issue the "last will" publication (if any). This will be received by other clients subscribed to the topic.
If the client determines that connectivity has been lost it will close the socket and periodically attempt to reconnect until it succeeds.
In the event of failing connectivity client and server publications with qos == 0 may be lost. The behaviour of qos == 1 packets is described below.
These behave as follows. The client waits for response_time. If no
acknowledgment has been received it re-publishes it, up to MAX_REPUBS times.
In the absence of acknowledgment the network is presumed to be down. The client
reconnects as described above. The publication is then attempted again as a new
message with a different PID. (The new PID proved necessary for Mosquitto to
recognise the message).
This effectively guarantees the reception of a qos == 1 publication, with the proviso that the publishing coroutine will block until reception has been acknowledged.
It is permissible for qos == 1 publications to run concurrently with each paused pending acknowledgement, however this has implications for resource constrained devices. See Section 4.4.
Where the client is subscribed to a topic with qos == 1 and a publication with
qos == 1 occurs the broker will re-publish until an acknowledgment is
received. If the broker deems that connectivity has failed it waits for the
client to reconnect. If the client was configured with clean set True,
qos == 1 messages published during the outage will be lost. Otherwise they will
be received in quick succession (which can overflow the buffer on an ESP8266
resulting in LmacRxBlk:1 messages).
The module allows concurrent publications and registration of subscriptions.
When using qos == 1 publications on hardware with limited resources such as ESP8266 it is wise to avoid concurrency by implementing a single publication task. In such cases if a publication queue is required it should be implemented by the application.
On capable hardware it is valid to have multiple coroutines performing qos == 1 publications asynchronously, but there are implications where connectivity with the broker is slow: an accumulation of tasks waiting on PUBACK packets implies consumption of resources.
The WiFi and Connect coroutines should run to completion quickly relative to
the time required to connect and disconnect from the network. Aim for 2 seconds
maximum. Alternatively the Connect coro can run indefinitely so long as it
terminates if the isconnected() method returns False.
The subscription callback will block publications and the reception of further subscribed messages and should therefore be designed for a fast return.
A contributor (Kevin Köck) was concerned that, in the case of a connectivity outage, a publication might be delayed to the point where it was excessively outdated. He wanted to implement a timeout to cancel the publication if an outage caused high latency. This is arguably a limitation of MQTT3.1.1 - please see MQTTv5 Support.
The following notes are a discussion of workrounds for V3.1.1.
Simple cancellation of a publication task is not recommended because it can disrupt the MQTT protocol. There are several ways to address this:
MQTTClient and acquire the self.lock object before issuing
the cancellation. The self.lock object protects a protocol sequence so that
it cannot be disrupted by another task. This was the method successfully
adopted and can be seen in mqtt_as_timeout.py.This was not included in the library mainly because most use cases are covered by use of a timestamp. Other reasons are documented in the code comments.
The library aims to handle connectivity outages transparently, however power
cycling of the client must be considered at application level. When the
application calls the client's connect method any failure will cause an
OSError to be raised. This is by design because the action to be taken is
application-dependent. A check on WiFi or broker function may be required.
There may be a need to fall back to a different network. In other applications
brief power outages may be expected: when power resumes the client will simply
reconnect. If an error occurs the application might wait for a period before
re-trying.
When that initial connection has been achieved, subsequent connections caused by network outages are handled transparently to the application.
The behaviour of "clean session" should be considered in this context. If the
clean flag is False and a long power outage occurs there may be a large
backlog of messages. This can cause problems on resource constrained clients,
notably if the client has been taken out of service for a few days. This module
addresses this by enabling behaviour which differs between the power up case
and the case of a network outage.
The clean_init flag determines behaviour on power up, while clean defines
behaviour after a connectivity outage. If clean_init is True and clean is
False, on power up prior session state is discarded. The client reconnects
with clean==False. It reconnects similarly after connectivity outages. Hence,
after power up, subscriptions will meet the qos==1 guarantee for messages
published during connectivity outages.
If both flags are False normal non-clean behaviour ensues with the potential
for substantial backlogs after long power outages.
If on power up both flags are True the broker will discard session state
during connectivity (and hence power) outages. This implies a loss of messages
published during connectivity outages(MQTT spec 3.1.2.4 Clean Session).
Also discussed here.
Version 0.8.2 introduced an optimisation whereby incoming messages are read into a pre-allocated buffer. This avoids allocation and improves performance. The change was done in a way that avoids breaking existing code. Allocation may be further reduced by setting two module variables. These are (with defaults):
IBUFSIZE = 50
MSG_BYTES = True
Any changes should be made before instantiating the client, e.g.:
import mqtt_as
mqtt_as.IBUFSIZE = 5_000
client = MQTTClient(config)Socket reads are into a pre-allocated buffer. If a message arrives which is too large, the buffer is extended to accept it. This implies allocation. Consider a case where a long message arrives after a long period where only short messages are received. By this time the RAM may have become fragmented, making the large allocation fail. If it is known that large messages may arrive, setting a large buffer size at the outset - prior to fragmentation - will avoid this problem.
By default, incoming messages are copied before being made available to the
application. This implies allocation. It is done to ensure message integrity
under all conditions. If the event interface is used, copying occurs regardless
of MSG_BYTES.
In the case of the callback interface where MSG_BYTES is False, a
memoryview of the buffer is passed to the callback, avoiding allocation.
The following code is safe where a memoryview is returned:
# Subscription callback
def sub_cb(topic, msg, retained):
# Synchronous code handles the messageHowever this presents a hazard if a memoryview is returned:
# Subscription callback
def sub_cb(topic, msg, retained):
asyncio.create_task(process_message(topic, msg))A fault arises if another message arrives before process_message is complete.
The buffer contents will change, causing corruption.
Normal operation of mqtt_as is based on attempting to keep the link up as
much as possible. This assures minimum latency for subscriptions but implies
power draw. The machine module supports two power saving modes: lightsleep
and deepsleep. Currently asyncio supports neither of these modes. The
notes below may be relevant to any application which deliberately closes and
re-opens the link to the broker.
Maximum power savings may be achieved by periodically connecting, handling
publications and pending subscriptions, and entering deepsleep. With suitable
hardware it is possible to produce an MQTT client with very low average power
consumption. This is done by keeping the application run time short and using
machine.deepsleep to sleep for a period. When the period expires the board
resets and main.py re-starts the application.
Hardware tested was the UM Feather S2 available from Adafruit. My sample consumes only 66μA in deepsleep mode. It has a switchable LDO regulator allowing external sensors to be powered down when the host is in deepsleep. It also supports battery operation via a LiPo cell with USB charging. A Pyboard D with WBUS-DIP28 has similar properties.
The test script lptest_min.py wakes up periodically and connects to WiFi. It publishes the value from the onboard light sensor, and subscribes to the topic "foo_topic". Any matching publications which occurred during deepsleep are received and revealed by flashing the blue LED.
Note that deepsleep disables USB. This is inconvenient in development. The
script has a test mode in which deepsleep is replaced by time.sleep and
machine.soft_reset keeping the USB link active. An alternative approach to
debugging is to use a UART with an FTDI adaptor. Such a link can survive a
deep sleep.
Each time the client goes into deepsleep it issues .disconnect(). This sends
an MQTT DISCONNECT packet to the broker suppressing the last will as per MQTT
spec para 3.1.2.5. The reasoning is that deepsleep periods are likely to be
much longer than the keepalive time. Using .disconnect() ensures that a last
will message is only triggered in the event of a failure such as a program
crash.
In applications which close the connection and deepsleep, power consumption may
be further reduced by setting the quick arg to .connect. On connecting or
re-connecting after an outage a check is made to ensure that WiFi connectivity
is stable. Quick connection skips this check on initial connection only, saving
several seconds. The reasoning here is that any error in initial connection
must be handled by the application. The test script sleeps for retry seconds
before re-trying the connection.
The library is not designed for use in cases where the system goes into
lightsleep. Firstly asyncio does not support lightsleep on all platforms -
notably on STM where the ticks_ms clock (crucial to task scheduling) stops
for the duration of lightsleep.
Secondly the library has no mechanism to ensure all tasks are shut down cleanly
after issuing .disconnect. This calls into question any application that
issues .disconnect and then attempts to reconnect. This issue does not arise
with deepsleep because the host effectively powers down. When the sleep
ends, asyncio and necessary tasks start as in a power up event.
These problems have been resolved by users for specific applications with forks
of the library. Given the limitations of asyncio I do not plan to write a
general solution.
This document
describes an MQTT client for ESP32 or ESP8266 which uses ESPNOw to communicate
with a gateway running mqtt_as. The client does not need to connect to WiFi
each time it wakes, saving power. The gateway can be shared between multiple
clients.
Drawbacks are the need for an always-on gateway, and the fact that only a subset of MQTT V3.1.1 capabilities is supported.
mqtt introduction
mosquitto server
mosquitto client publish
mosquitto client subscribe
MQTT 3.1.1 spec
MQTTv5 spec
python client for PC's
Unofficial MQTT FAQ
List of public brokers
On the initial connection attempt the broker may reject the attempt. In this
instance an OSError will be raised showing two numbers. The first number
should be 0x2002 which is the MQTT CONNACK fixed header. The second
is CONNACK variable header byte 2 which indicates the reason for failure as
follows:
| Value | Reason |
|---|---|
| 1 | Unacceptable protocol version. |
| 2 | Client identifier rejected. |
| 3 | MQTT service unavailable. |
| 4 | Username or password have an invalid format. |
| 5 | Client is not authorised to connect. |
See MQTT spec section 3.2.2.
The Hive MQ site offers a free web-based broker which is more secure than public brokers. With a public broker anyone can detect and subscribe to your publications. Hive MQ gives you a unique broker internet address which requires a password to access. TLS is mandatory but does not require certificates.
A simple GitHub registration gets you:
Typical usage:
config['user'] = 'my_username'
config['password'] = 'my_password'
broker = 'unique broker address' # e.g long_hex_string.s2.eu.hivemq.cloud
config['server'] = broker
config['ssl'] = True
config['ssl_params'] = {"server_hostname": broker}The free service is scalable (at cost) to large commercial deployments.
The following are the allowable keys:
According to this post the following platforms use mbedtls:
See this post
for details of how to use client certificates with a mosquitto broker.
Note also that TLS with client certificates requires the cliet's clock to be
approximately correct. This can be achieved with an NTP query. If mosquitto
is run on a local server it also runs the NTP daemon. A high availability
option is to run the NTP query against the local server. See
this doc,
also the official ntptime module.
See this link for information on creating client certificates and a Bash script for doing so.