Insertion of thousands of contact entries using applyBatch is slow

Question

I'm developing an application where I need to insert lots of Contact entries. At the current time approx 600 contacts with a total of 6000 phone numbers. The biggest contact has 1800 phone numbers.

Status as of today is that I have created a custom Account to hold the Contacts, so the user can select to see the contact in the Contacts view.

But the insertion of the contacts is painfully slow. I insert the contacts using ContentResolver.applyBatch. I've tried with different sizes of the ContentProviderOperation list(100, 200, 400), but the total running time is approx. the same. To insert all the contacts and numbers takes about 30 minutes!

Most issues I've found regarding slow insertion in SQlite brings up transactions. But since I use the ContentResolver.applyBatch-method I don't control this, and I would assume that the ContentResolver takes care of transaction management for me.

So, to my question: Am I doing something wrong, or is there anything I can do to speed this up?

Anders

Edit: @jcwenger: Oh, I see. Good explanation!

So then I will have to first insert into the raw_contacts table, and then the datatable with the name and numbers. What I'll lose is the back reference to the raw_id which I use in the applyBatch.

So I'll have to get all the id's of the newly inserted raw_contacts rows to use as foreign keys in the data table?

Answer 1

Use ContentResolver.bulkInsert (Uri url, ContentValues[] values) instead of ApplyBatch()

ApplyBatch (1) uses transactions and (2) it locks the ContentProvider once for the whole batch instead locking/unlocking once per operation. because of this, it is slightly faster than doing them one at a time (non-batched).

However, since each Operation in the Batch can have a different URI and so on, there's a huge amount of overhead. "Oh, a new operation! I wonder what table it goes in... Here, I'll insert a single row... Oh, a new operation! I wonder what table it goes in..." ad infinitium. Since most of the work of turning URIs into tables involves lots of string comparisons, it's obviously very slow.

By contrast, bulkInsert applies a whole pile of values to the same table. It goes, "Bulk insert... find the table, okay, insert! insert! insert! insert! insert!" Much faster.

It will, of course, require your ContentResolver to implement bulkInsert efficiently. Most do, unless you wrote it yourself, in which case it will take a bit of coding.

Answer 2

bulkInsert: For those interested, here is the code that I was able to experiment with. Pay attention to how we can avoid some allocations for int/long/floats :) this could save more time.

private int doBulkInsertOptimised(Uri uri, ContentValues values[]) {
    long startTime = System.currentTimeMillis();
    long endTime = 0;
    //TimingInfo timingInfo = new TimingInfo(startTime);

    SQLiteDatabase db = mOpenHelper.getWritableDatabase();

    DatabaseUtils.InsertHelper inserter =
        new DatabaseUtils.InsertHelper(db, Tables.GUYS); 

    // Get the numeric indexes for each of the columns that we're updating
    final int guiStrColumn = inserter.getColumnIndex(Guys.STRINGCOLUMNTYPE);
    final int guyDoubleColumn = inserter.getColumnIndex(Guys.DOUBLECOLUMNTYPE);
//...
    final int guyIntColumn = inserter.getColumnIndex(Guys.INTEGERCOLUMUNTYPE);

    db.beginTransaction();
    int numInserted = 0;
    try {
        int len = values.length;
        for (int i = 0; i < len; i++) {
            inserter.prepareForInsert();

            String guyID = (String)(values[i].get(Guys.GUY_ID)); 
            inserter.bind(guiStrColumn, guyID);


            // convert to double ourselves to save an allocation.
            double d = ((Number)(values[i].get(Guys.DOUBLECOLUMNTYPE))).doubleValue();
            inserter.bind(guyDoubleColumn, lat);


            // getting the raw Object and converting it int ourselves saves
            // an allocation (the alternative is ContentValues.getAsInt, which
            // returns a Integer object)

            int status = ((Number) values[i].get(Guys.INTEGERCOLUMUNTYPE)).intValue();
            inserter.bind(guyIntColumn, status);

            inserter.execute();
        }
        numInserted = len;
        db.setTransactionSuccessful();
    } finally {
        db.endTransaction();
        inserter.close();

        endTime = System.currentTimeMillis();

        if (LOGV) {
            long timeTaken = (endTime - startTime);
            Log.v(TAG, "Time taken to insert " + values.length + " records was " + timeTaken + 
                    " milliseconds " + " or " + (timeTaken/1000) + "seconds");
        }
    }
    getContext().getContentResolver().notifyChange(uri, null);
    return numInserted;
}

Answer 3

An example of on how to override the bulkInsert(), in order to speed up multiples insert, can be found here

Answer 4

@jcwenger At first, after read your post, I think that's the reason of bulkInsert is quicker than ApplyBatch, but after read the code of Contact Provider, I don't think so. 1.You said ApplyBatch use transactions, yes, but bulkInsert also use transactions. Here is the code of it:

public int bulkInsert(Uri uri, ContentValues[] values) {
    int numValues = values.length;
    mDb = mOpenHelper.getWritableDatabase();
    mDb.beginTransactionWithListener(this);
    try {
        for (int i = 0; i < numValues; i++) {
            Uri result = insertInTransaction(uri, values[i]);
            if (result != null) {
                mNotifyChange = true;
            }
            mDb.yieldIfContendedSafely();
        }
        mDb.setTransactionSuccessful();
    } finally {
        mDb.endTransaction();
    }
    onEndTransaction();
    return numValues;
}

That is to say, bulkInsert also use transations.So I don't think that's the reason. 2.You said bulkInsert applies a whole pile of values to the same table.I'm sorry I can't find related code in the source code of froyo.And I want to know how could you find that?Could you tell me?

The reason I think is that:

bulkInsert use mDb.yieldIfContendedSafely() while applyBatch use mDb.yieldIfContendedSafely(SLEEP_AFTER_YIELD_DELAY)/*SLEEP_AFTER_YIELD_DELAY = 4000*/

after reading the code of SQLiteDatabase.java, I find that, if set a time in yieldIfContendedSafely, it will do a sleep, but if you don't set the time, it will not sleep.You can refer to the code below which is a piece of code of SQLiteDatabase.java

private boolean yieldIfContendedHelper(boolean checkFullyYielded, long     sleepAfterYieldDelay) {
    if (mLock.getQueueLength() == 0) {
        // Reset the lock acquire time since we know that the thread was willing to yield
        // the lock at this time.
        mLockAcquiredWallTime = SystemClock.elapsedRealtime();
        mLockAcquiredThreadTime = Debug.threadCpuTimeNanos();
        return false;
    }
    setTransactionSuccessful();
    SQLiteTransactionListener transactionListener = mTransactionListener;
    endTransaction();
    if (checkFullyYielded) {
        if (this.isDbLockedByCurrentThread()) {
            throw new IllegalStateException(
                    "Db locked more than once. yielfIfContended cannot yield");
        }
    }
    if (sleepAfterYieldDelay > 0) {
        // Sleep for up to sleepAfterYieldDelay milliseconds, waking up periodically to
        // check if anyone is using the database.  If the database is not contended,
        // retake the lock and return.
        long remainingDelay = sleepAfterYieldDelay;
        while (remainingDelay > 0) {
            try {
                Thread.sleep(remainingDelay < SLEEP_AFTER_YIELD_QUANTUM ?
                        remainingDelay : SLEEP_AFTER_YIELD_QUANTUM);
            } catch (InterruptedException e) {
                Thread.interrupted();
            }
            remainingDelay -= SLEEP_AFTER_YIELD_QUANTUM;
            if (mLock.getQueueLength() == 0) {
                break;
            }
        }
    }
    beginTransactionWithListener(transactionListener);
    return true;
}

I think that's the reason of bulkInsert is quicker than applyBatch.

Any question please contact me.

Answer 5

I get the basic solution for you, use "yield points" in batch operation.

The flip side of using batched operations is that a large batch may lock up the database for a long time preventing other applications from accessing data and potentially causing ANRs ("Application Not Responding" dialogs.)

To avoid such lockups of the database, make sure to insert "yield points" in the batch. A yield point indicates to the content provider that before executing the next operation it can commit the changes that have already been made, yield to other requests, open another transaction and continue processing operations.

A yield point will not automatically commit the transaction, but only if there is another request waiting on the database. Normally a sync adapter should insert a yield point at the beginning of each raw contact operation sequence in the batch. See withYieldAllowed(boolean).

I hope it's may be useful for you.