Introduction to Multi-Document ACID Transactions in Python

#Introduction

Multi-document transactions arrived in MongoDB 4.0 in June 2018. MongoDB has always been transactional around updates to a single document. Now, with multi-document ACID transactions we can wrap a set of database operations inside a start and commit transaction call. This ensures that even with inserts and/or updates happening across multiple collections and/or databases, the external view of the data meets ACID constraints.

To demonstrate transactions in the wild we use a trivial example app that emulates a flight booking for an online airline application. In this simplified booking we need to undertake three operations:

• Allocate a seat in the seat_collection
• Pay for the seat in the payment_collection
• Update the count of allocated seats and sales in the audit_collection

For this application we will use three separate collections for these documents as detailed above. The code in transactions_main.py updates these collections in serial unless the --usetxns argument is used. We then wrap the complete set of operations inside an ACID transaction. The code in transactions_main.py is built directly using the MongoDB Python driver (Pymongo 3.7.1).

The goal of this code is to demonstrate to the Python developers just how easy it is to covert existing code to transactions if required or to port older SQL based systems.

#Setting up your environment

The following files can be found in the associated github repo, pymongo-transactions.

• gitignore : Standard Github .gitignore for Python.
• LICENSE : Apache's 2.0 (standard Github) license.
• Makefile : Makefile with targets for default operations.
• transaction_main.py : Run a set of writes with and without transactions. Run python transactions_main.py -h for help.
• transactions_retry.py : The file containing the transactions retry functions.
• watch_transactions.py : Use a MongoDB change stream to watch collections as they change when transactions_main.py is running.
• kill_primary.py : Starts a MongoDB replica set (on port 7100) and kills the primary on a regular basis. This is used to emulate an election happening in the middle of a transaction.
• featurecompatibility.py : check and/or set feature compatibility for the database (it needs to be set to "4.0" for transactions).

You can clone this repo and work alongside us during this blog post (please file any problems on the Issues tab in Github).

We assume for all that follows that you have Python 3.6 or greater correctly installed and on your path.

The Makefile outlines the operations that are required to setup the test environment.

All the programs in this example use a port range starting at 27100 to ensure that this example does not clash with an existing MongoDB installation.

#Preparation

To setup the environment you can run through the following steps manually. People that have make can speed up installation by using the make install command.

1
$ cd pymongo-transactions
2
$ virtualenv -p python3 venv
3
$ source venv/bin/activate

1
pip install --upgrade pymongo

1
pip install mtools

1
pip install psutil

1
mlaunch init --port 27100 --replicaset --name "txntest"

1
$ cd pymongo-transactions
2
$ make

#Running the transactions example

The transactions example consists of two python programs.

• transaction_main.py,
• watch_transactions.py.

1
$ python transaction_main.py -h
2
usage: transaction_main.py [-h] [--host HOST] [--usetxns] [--delay DELAY]
3
                           [--iterations ITERATIONS]
4
                           [--randdelay RANDDELAY RANDDELAY]
5
6
optional arguments:
7
  -h, --help            show this help message and exit
8
  --host HOST           MongoDB URI [default: mongodb://localhost:27100,localh
9
                        ost:27101,localhost:27102/?replicaSet=txntest&retryWri
10
                        tes=true]
11
  --usetxns             Use transactions [default: False]
12
  --delay DELAY         Delay between two insertion events [default: 1.0]
13
  --iterations ITERATIONS
14
                        Run N iterations. O means run forever
15
  --randdelay RANDDELAY RANDDELAY
16
                        Create a delay set randomly between the two bounds
17
                        [default: None]

1
$ python transaction_main.py
2
using collection: SEATSDB.seats
3
using collection: PAYMENTSDB.payments
4
using collection: AUDITDB.audit
5
Using a fixed delay of 1.0
6
7
1. Booking seat: '1A'
8
1. Sleeping: 1.000
9
1. Paying 330 for seat '1A'
10
2. Booking seat: '2A'
11
2. Sleeping: 1.000
12
2. Paying 450 for seat '2A'
13
3. Booking seat: '3A'
14
3. Sleeping: 1.000
15
3. Paying 490 for seat '3A'
16
4. Booking seat: '4A'
17
4. Sleeping: 1.000
18
^C

1
Traceback (most recent call last):
2
  File "transaction_main.py", line 175, in
3
    total_delay = total_delay + run_transaction_with_retry( booking_functor, session)
4
  File "/Users/jdrumgoole/GIT/pymongo-transactions/transaction_retry.py", line 52, in run_transaction_with_retry
5
    with session.start_transaction():
6
AttributeError: 'ClientSession' object has no attribute 'start_transaction'

#Watching Transactions

To actually see the effect of transactions we need to watch what is happening inside the collections SEATSDB.seats and PAYMENTSDB.payments.

We can do this with watch_transactions.py. This script uses MongoDB Change Streams to see what's happening inside a collection in real-time. We need to run two of these in parallel so it's best to line them up side by side.

Here is the watch_transactions.py program:

1
$ python watch_transactions.py -h
2
usage: watch_transactions.py [-h] [--host HOST] [--collection COLLECTION]
3
4
optional arguments:
5
  -h, --help            show this help message and exit
6
  --host HOST           mongodb URI for connecting to server [default:
7
                        mongodb://localhost:27100/?replicaSet=txntest]
8
  --collection COLLECTION
9
                        Watch  [default:
10
                        PYTHON_TXNS_EXAMPLE.seats_collection]

We need to watch each collection so in two separate terminal windows start the watcher.

Window 1:

1
$ python watch_transactions.py --watch seats
2
Watching: seats
3
...

Window 2:

1
$ python watch_transactions.py --watch payments
2
Watching: payments
3
...

#What happens when you run without transactions?

Lets run the code without transactions first. If you examine the transaction_main.py code you will see a function book_seats.

1
def book_seat(seats, payments, audit, seat_no, delay_range, session=None):
2
    '''
    Run two inserts in sequence.
    If session is not None we are in a transaction

    :param seats: seats collection
    :param payments: payments collection
    :param seat_no: the number of the seat to be booked (defaults to row A)
    :param delay_range: A tuple indicating a random delay between two ranges or a single float fixed delay
    :param session: Session object required by a MongoDB transaction
    :return: the delay_period for this transaction
    '''
3
    price = random.randrange(200, 500, 10)
4
    if type(delay_range) == tuple:
5
        delay_period = random.uniform(delay_range[0], delay_range[1])
6
    else:
7
        delay_period = delay_range
8
9
    # Book Seat
10
    seat_str = "{}A".format(seat_no)
11
    print(count( i, "Booking seat: '{}'".format(seat_str)))
12
    seats.insert_one({"flight_no" : "EI178",
13
                      "seat"      : seat_str,
14
                      "date"      : datetime.datetime.utcnow()},
15
                     session=session)
16
    print(count( seat_no, "Sleeping: {:02.3f}".format(delay_period)))
17
    #pay for seat
18
    time.sleep(delay_period)
19
    payments.insert_one({"flight_no" : "EI178",
20
                         "seat"      : seat_str,
21
                         "date"      : datetime.datetime.utcnow(),
22
                         "price"     : price},
23
                        session=session)
24
    audit.update_one({ "audit" : "seats"}, { "$inc" : { "count" : 1}}, upsert=True)
25
    print(count(seat_no, "Paying {} for seat '{}'".format(price, seat_str)))
26
27
    return delay_period

This program emulates a very simplified airline booking with a seat being allocated and then paid for. These are often separated by a reasonable time frame (e.g. seat allocation vs external credit card validation and anti-fraud check) and we emulate this by inserting a delay. The default is 1 second.

Now with the two watch_transactions.py scripts running for seats_collection and payments_collection we can run transactions_main.py as follows:

1
$ python transaction_main.py

The first run is with no transactions enabled.

The bottom window shows transactions_main.py running. On the top left we are watching the inserts to the seats collection. On the top right we are watching inserts to the payments collection.

We can see that the payments window lags the seats window as the watchers only update when the insert is complete. Thus seats sold cannot be easily reconciled with corresponding payments. If after the third seat has been booked we CTRL-C the program we can see that the program exits before writing the payment. This is reflected in the Change Stream for the payments collection which only shows payments for seat 1A and 2A versus seat allocations for 1A, 2A and 3A.

If we want payments and seats to be instantly reconcilable and consistent we must execute the inserts inside a transaction.

#What happens when you run with Transactions?

Now lets run the same system with --usetxns enabled.

1
$ python transaction_main.py --usetxns

We run with the exact same setup but now set --usetxns.

Note now how the change streams are interlocked and are updated in parallel. This is because all the updates only become visible when the transaction is committed. Note how we aborted the third transaction by hitting CTRL-C. Now neither the seat nor the payment appear in the change streams unlike the first example where the seat went through.

This is where transactions shine in world where all or nothing is the watchword. We never want to keeps seats allocated unless they are paid for.

#What happens during failure?

In a MongoDB replica set all writes are directed to the Primary node. If the primary node fails or becomes inaccessible (e.g. due to a network partition) writes in flight may fail. In a non-transactional scenario the driver will recover from a single failure and retry the write. In a multi-document transaction we must recover and retry in the event of these kinds of transient failures. This code is encapsulated in transaction_retry.py. We both retry the transaction and retry the commit to handle scenarios where the primary fails within the transaction and/or the commit operation.

1
def commit_with_retry(session):
2
    while True:
3
        try:
4
            # Commit uses write concern set at transaction start.
5
            session.commit_transaction()
6
            print("Transaction committed.")
7
            break
8
        except (pymongo.errors.ConnectionFailure, pymongo.errors.OperationFailure) as exc:
9
            # Can retry commit
10
            if exc.has_error_label("UnknownTransactionCommitResult"):
11
                print("UnknownTransactionCommitResult, retrying "
12
                      "commit operation ...")
13
                continue
14
            else:
15
                print("Error during commit ...")
16
                raise
17
18
def run_transaction_with_retry(functor, session):
19
    assert (isinstance(functor, Transaction_Functor))
20
    while True:
21
        try:
22
            with session.start_transaction():
23
                result=functor(session)  # performs transaction
24
                commit_with_retry(session)
25
            break
26
        except (pymongo.errors.ConnectionFailure, pymongo.errors.OperationFailure) as exc:
27
            # If transient error, retry the whole transaction
28
            if exc.has_error_label("TransientTransactionError"):
29
                print("TransientTransactionError, retrying "
30
                      "transaction ...")
31
                continue
32
            else:
33
                raise
34
35
    return result

In order to observe what happens during elections we can use the script kill_primary.py. This script will start a replica-set and continuously kill the primary.

1
$ make kill_primary
2
. venv/bin/activate && python kill_primary.py
3
no nodes started.
4
Current electionTimeoutMillis: 500
5
1. (Re)starting replica-set
6
no nodes started.
7
1. Getting list of mongod processes
8
Process list written to mlaunch.procs
9
1. Getting replica set status
10
1. Killing primary node: 31029
11
1. Sleeping: 1.0
12
2. (Re)starting replica-set
13
launching: "/usr/local/mongodb/bin/mongod" on port 27101
14
2. Getting list of mongod processes
15
Process list written to mlaunch.procs
16
2. Getting replica set status
17
2. Killing primary node: 31045
18
2. Sleeping: 1.0
19
3. (Re)starting replica-set
20
launching: "/usr/local/mongodb/bin/mongod" on port 27102
21
3. Getting list of mongod processes
22
Process list written to mlaunch.procs
23
3. Getting replica set status
24
3. Killing primary node: 31137
25
3. Sleeping: 1.0

kill_primary.py resets electionTimeOutMillis to 500ms from its default of 10000ms (10 seconds). This allows elections to resolve more quickly for the purposes of this test as we are running everything locally.

Once kill_primary.py is running we can start up transactions_main.py again using the --usetxns argument.

1
$ make usetxns
2
. venv/bin/activate && python transaction_main.py --usetxns
3
Forcing collection creation (you can't create collections inside a txn)
4
Collections created
5
using collection: PYTHON_TXNS_EXAMPLE.seats
6
using collection: PYTHON_TXNS_EXAMPLE.payments
7
using collection: PYTHON_TXNS_EXAMPLE.audit
8
Using a fixed delay of 1.0
9
Using transactions
10
11
1. Booking seat: '1A'
12
1. Sleeping: 1.000
13
1. Paying 440 for seat '1A'
14
Transaction committed.
15
2. Booking seat: '2A'
16
2. Sleeping: 1.000
17
2. Paying 330 for seat '2A'
18
Transaction committed.
19
3. Booking seat: '3A'
20
3. Sleeping: 1.000
21
TransientTransactionError, retrying transaction ...
22
3. Booking seat: '3A'
23
3. Sleeping: 1.000
24
3. Paying 240 for seat '3A'
25
Transaction committed.
26
4. Booking seat: '4A'
27
4. Sleeping: 1.000
28
4. Paying 410 for seat '4A'
29
Transaction committed.
30
5. Booking seat: '5A'
31
5. Sleeping: 1.000
32
5. Paying 260 for seat '5A'
33
Transaction committed.
34
6. Booking seat: '6A'
35
6. Sleeping: 1.000
36
TransientTransactionError, retrying transaction ...
37
6. Booking seat: '6A'
38
6. Sleeping: 1.000
39
6. Paying 380 for seat '6A'
40
Transaction committed.
41
...

As you can see during elections the transaction will be aborted and must be retried. If you look at the transaction_rety.py code you will see how this happens. If a write operation encounters an error it will throw one of the following exceptions:

• pymongo.errors.ConnectionFailure
• pymongo.errors.OperationFailure

Within these exceptions there will be a label called TransientTransactionError. This label can be detected using the has_error_label(label) function which is available in pymongo 3.7.x. Transient errors can be recovered from and the retry code in transactions_retry.py has code that retries for both writes and commits (see above).

#Conclusion

Multi-document transactions are the final piece of the jigsaw for SQL developers who have been shying away from trying MongoDB. ACID transactions make the programmer's job easier and give teams that are migrating from an existing SQL schema a much more consistent and convenient transition path.

As most migrations involving a move from highly normalised data structures to more natural and flexible nested JSON documents one would expect that the number of required multi-document transactions will be less in a properly constructed MongoDB application. But where multi-document transactions are required programmers can now include them using very similar syntax to SQL.

With ACID transactions in MongoDB 4.0 it can now be the first choice for an even broader range of application use cases.

To try it locally download MongoDB 4.0.