How the File Upload Process Works

The File Migration service maintains in memory information regarding directories, subdirectories, and files within each directory tree to monitor. When started, the service first looks for changed and new files in the configured directories. To detect changes to existing files, the service looks at the extended attribute user.s3uploadtime, which is set on each file and contains the modification time (in milliseconds) when the file was last uploaded to the AWS S3 bucket. It then places the new and changed files in a queue for upload, sorted by the oldest modification timestamp. The service then uploads multiple files in parallel and provides logs as files upload, fail to upload, or complete uploading.

Note: For each directory tree, you can set up any file or directory that matches a specified regex pattern to be ignored. For example, all files ending in .tmp or all files in /tmp directory can be set up to be ignored. Additionally, only files are considered for upload; empty directories, symbolic links, and tables or streams are ignored.

Scanning for New and Changed Files

The service scans for new and changed files in the monitored directories using a full scan or a lite scan.

The service performs a full scan when it is first started and periodically or randomly for a subset of the directories in the directory tree after. This allows you to spread the cost of the full scan over time rather than performing a full scan of all directories at once.

During a full scan, the service examines each file in a directory to determine if it has changed since the last time the directory was examined. If the file contains changes, then the file is examined more carefully.

If the file was uploaded, the service retrieves the modification time of the file, which was set on the file at the time the file was uploaded. If the file was not uploaded or if the file was uploaded but the modification time changed since the last upload, the service places the file in the queue for upload.

A full scan also discovers new directories and adds them to the list of directories to be monitored.

The service performs an incremental scan to detect the last modified timestamp on each directory recursively using the in-memory directory tree from the last scan. If there is no change in the timestamp since the last scan, the service does not examine the directory contents and proceeds to the next directory; this makes the incremental scan very fast. If there is a change in the timestamp since the last scan, the service performs a full scan on the directory.
Note: Before starting an incremental scan, the service checks to see if a directory tree is due for a full scan. Only a full scan allows the service to detect if a file was modified after upload. With a lite scan, the service checks only the modification time of the directory, which is not changed by the modification (such as due to late writes) of a file.

Queueing Files for Upload

Files are placed in the queue with a configurable delay from the last known modification time of the file to avoid uploading a file that is still changing.

For example, with the default of 15 seconds, a file that was modified 5 seconds ago would remain on the queue for another 10 seconds, whereas a file modified 5 minutes ago would be immediately considered for upload. This allows files to be uploaded roughly in order by last modification time.

The service places (by default, 10) files in the queue for parallel upload. The scanning service is paused if the pending queue exceeds 10 times the number of files allowed for parallel update to avoid too many pending files.

Uploading Files to S3

Uploads are in one of three states: waiting to be uploaded, uploading, finished with uploading. By default, the service performs a check on the status of the files being uploaded every 3 seconds and detects uploads that are approaching completion. The service then sets the extended attribute (user.s3uploadtime) to the file on MapR Filesystem to indicate that the file was uploaded.

If an error occurs during upload, the file is requeued for another try after a brief wait, which is 5 minutes by default. If the file could not be uploaded after a specified number of attempts, the file is dequeued to allow other files to upload. Note that the file is rediscovered, if it is still there, during the next full scan and queued for upload.

Note: When files are uploaded, although the file name remains the same, but the leading slash is removed and the full path is converted to a key on the AWS S3 bucket. For example, /tmp/foo after upload becomes the key tmp/foo on AWS S3 bucket.

Setting Extended Attributes

When the upload completes, the service uses an extended attribute, user.s3uploadtime, to set the timestamp on the file to when the file was uploaded. This extended attribute, which contains the last upload timestamp, is used by the service to determine if a file has changed since the file was last uploaded.

Purging the Files and Directories

The service also optionally deletes files after uploading and deletes empty directories. You can configure file deletion after upload in hours or fractions of an hour and configure empty directory deletion in minutes. Both can be set to zero, which disables the deletion function. Both functions are disabled by default.

Deletion of files and directories happen only during a full scan:

For files, the modification timestamp on the file is examined to see how far back in time it was last modified.

If the file has been uploaded in its present form and if it matches or exceeds the interval configured for file deletion, the file is immediately deleted. If it does not match the configured interval for file deletion, a message is printed (to the log file, filemigrate.log) as to when the file will be deleted if it will be deleted fairly soon.

Files that were never uploaded (such as the files that were ignored) will not be deleted by the service.

For directories, the contents of the directory and the modification timestamp on the directory are examined.

If the directory is empty and the modification timestamp matches or exceeds the configured interval of time for directory deletion, the directory is deleted.

The act of deleting a file or directory from a parent directory causes the modification timestamp of the parent directory to change. Therefore, an idle or empty (parent) directory is deleted only after the configured interval of time since the deletion of its children (file or subdirectory).

The recursive deletion of directories has a slow bottom up behavior: first empty leaf directories are deleted and then the parents are deleted during a future full scan after the interval of time has elapsed.

If a directory contains files that were never uploaded (ignored), the directory will not be deleted.