common/chunkStatusLogger.go

// Copyright © 2017 Microsoft <wastore@microsoft.com>
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in
// all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
// THE SOFTWARE.

package common

import (
	"bufio"
	"fmt"
	"math"
	"os"
	"path"
	"sync/atomic"
	"time"
)

// Identifies a chunk. Always create with NewChunkID
type ChunkID struct {
	Name         string
	offsetInFile int64
	length       int64

	// What is this chunk's progress currently waiting on?
	// Must be a pointer, because the ChunkID itself is a struct.
	// When chunkID is passed around, copies are made,
	// but because this is a pointer, all will point to the same
	// value for waitReasonIndex (so when we change it, all will see the change)
	waitReasonIndex *int32

	// Like waitReasonIndex, but is effectively just a boolean to track whether we are done.
	// Must be a pointer, for same reason that waitReasonIndex is.
	// Can't be done just off waitReasonIndex because for downloads we actually
	// tell the jptm we are done before the chunk has been flushed out to disk, so
	// waitReasonIndex isn't yet ready to go to "Done" at that time.
	completionNotifiedToJptm *int32

	// TODO: it's a bit odd having two pointers in a struct like this.  Review, maybe we should always work
	//   with pointers to chunk ids, with nocopy?  If we do that, the two fields that are currently pointers
	//   can become non-pointers
	//   And maybe at that point, we would also put Length into chunkID, and use that in jptm.ReportChunkDone
}

func NewChunkID(name string, offsetInFile int64, length int64) ChunkID {
	dummyWaitReasonIndex := int32(0)
	zeroNotificationState := int32(0)
	return ChunkID{
		Name:                     name,
		offsetInFile:             offsetInFile,
		length:                   length,
		waitReasonIndex:          &dummyWaitReasonIndex, // must initialize, so don't get nil pointer on usage
		completionNotifiedToJptm: &zeroNotificationState,
	}
}

func NewPseudoChunkIDForWholeFile(name string) ChunkID {
	dummyWaitReasonIndex := int32(0)
	alreadyNotifiedNotificationState := int32(1) // so that these can never be notified to jptm's (doing so would be an error, because they are not real chunks)
	return ChunkID{
		Name:                     name,
		offsetInFile:             math.MinInt64,         // very negative, clearly not a real offset
		waitReasonIndex:          &dummyWaitReasonIndex, // must initialize, so don't get nil pointer on usage
		completionNotifiedToJptm: &alreadyNotifiedNotificationState,
	}
}

func (id ChunkID) SetCompletionNotificationSent() {
	if atomic.SwapInt32(id.completionNotifiedToJptm, 1) != 0 {
		panic("cannot complete the same chunk twice")
	}
}

func (id ChunkID) OffsetInFile() int64 {
	if id.offsetInFile < 0 {
		panic("Attempt to get negative file offset") // protects us against any mis-use of "pseudo chunks" with negative offsets
	}
	return id.offsetInFile
}

func (id ChunkID) IsPseudoChunk() bool {
	return id.offsetInFile < 0
}

func (id ChunkID) Length() int64 {
	return id.length
}

var EWaitReason = WaitReason{0, ""}

// WaitReason identifies the one thing that a given chunk is waiting on, at a given moment.
// Basically = state, phrased in terms of "the thing I'm waiting for"
type WaitReason struct {
	index int32
	Name  string
}

// Head (below) has index between GB and Body, just so the ordering is numerical ascending during typical chunk lifetime for both upload and download
// We use just the first letters of these when displaying perf states as we run (if enabled)
// so try to keep the first letters unique (except for Done and Cancelled, which are not displayed, and so may duplicate the first letter of something else)
func (WaitReason) Nothing() WaitReason              { return WaitReason{0, "Nothing"} }            // not waiting for anything
func (WaitReason) CreateLocalFile() WaitReason      { return WaitReason{1, "CreateLocalFile"} }    // creating the local file
func (WaitReason) RAMToSchedule() WaitReason        { return WaitReason{2, "RAM"} }                // waiting for enough RAM to schedule the chunk
func (WaitReason) WorkerGR() WaitReason             { return WaitReason{3, "Worker"} }             // waiting for a goroutine to start running our chunkfunc
func (WaitReason) FilePacer() WaitReason            { return WaitReason{4, "FilePacer"} }          // waiting until the file-level pacer says its OK to process another chunk
func (WaitReason) HeaderResponse() WaitReason       { return WaitReason{5, "Head"} }               // waiting to finish downloading the HEAD
func (WaitReason) Body() WaitReason                 { return WaitReason{6, "Body"} }               // waiting to finish sending/receiving the BODY
func (WaitReason) BodyReReadDueToMem() WaitReason   { return WaitReason{7, "BodyReRead-LowRam"} }  //waiting to re-read the body after a forced-retry due to low RAM
func (WaitReason) BodyReReadDueToSpeed() WaitReason { return WaitReason{8, "BodyReRead-TooSlow"} } // waiting to re-read the body after a forced-retry due to a slow chunk read (without low RAM)
func (WaitReason) Sorting() WaitReason              { return WaitReason{9, "Sorting"} }            // waiting for the writer routine, in chunkedFileWriter, to pick up this chunk and sort it into sequence
func (WaitReason) PriorChunk() WaitReason           { return WaitReason{10, "Prior"} }             // waiting on a prior chunk to arrive (before this one can be saved)
func (WaitReason) QueueToWrite() WaitReason         { return WaitReason{11, "Queue"} }             // prior chunk has arrived, but is not yet written out to disk
func (WaitReason) DiskIO() WaitReason               { return WaitReason{12, "DiskIO"} }            // waiting on disk read/write to complete
func (WaitReason) S2SCopyOnWire() WaitReason        { return WaitReason{13, "S2SCopyOnWire"} }     // waiting for S2S copy on wire get finished. extra status used only by S2S copy
func (WaitReason) Epilogue() WaitReason             { return WaitReason{14, "Epilogue"} }          // File-level epilogue processing (e.g. Commit block list, or other final operation on local or remote object (e.g. flush))

// extra ones for start of uploads (prior to chunk scheduling)
func (WaitReason) XferStart() WaitReason           { return WaitReason{15, "XferStart"} }
func (WaitReason) OpenLocalSource() WaitReason     { return WaitReason{16, "OpenLocalSource"} }
func (WaitReason) ModifiedTimeRefresh() WaitReason { return WaitReason{17, "ModifiedTimeRefresh"} }
func (WaitReason) LockDestination() WaitReason     { return WaitReason{18, "LockDestination"} }

func (WaitReason) ChunkDone() WaitReason { return WaitReason{19, "Done"} } // not waiting on anything. Chunk is done.
// NOTE: when adding new statuses please renumber to make Cancelled numerically the last, to avoid
// the need to also change numWaitReasons()
func (WaitReason) Cancelled() WaitReason { return WaitReason{20, "Cancelled"} } // transfer was cancelled.  All chunks end with either Done or Cancelled.

// TODO: consider change the above so that they don't create new struct on every call?  Is that necessary/useful?
//     Note: reason it's not using the normal enum approach, where it only has a number, is to try to optimize
//     the String method below, on the assumption that it will be called a lot.  Is that a premature optimization?

// Upload chunks go through these states, in this order.
// We record this set of states, in this order, so that when we are uploading GetCounts() can return
// counts for only those states that are relevant to upload (some are not relevant, so they are not in this list)
// AND so that GetCounts will return the counts in the order that the states actually happen when uploading.
// That makes it easy for end-users of the counts (i.e. logging and display code) to show the state counts
// in a meaningful left-to-right sequential order.
var uploadWaitReasons = []WaitReason{

	// pseudo-chunk stats (whole-of-file level) that happen before any chunks get scheduled
	EWaitReason.XferStart(),
	EWaitReason.OpenLocalSource(),
	EWaitReason.ModifiedTimeRefresh(),
	EWaitReason.LockDestination(),

	// These first two happen in the transfer initiation function (i.e. the chunkfunc creation loop)
	// So their total is constrained to the size of the goroutine pool that runs those functions.
	// (e.g. 64, given the GR pool sizing as at Feb 2019)
	EWaitReason.RAMToSchedule(),
	EWaitReason.DiskIO(),

	// This next one is used when waiting for a worker Go routine to pick up the scheduled chunk func.
	// Chunks in this state are effectively a queue of work waiting to be sent over the network
	EWaitReason.WorkerGR(),

	// Waiting until the per-file pacer (if any applies to this upload) says we can proceed
	EWaitReason.FilePacer(),

	// This is the actual network activity
	EWaitReason.Body(), // header is not separated out for uploads, so is implicitly included here

	EWaitReason.Epilogue(),
	// Plus Done/cancelled, which are not included here because not wanted for GetCounts
}

// Download chunks go through a larger set of states, due to needing to be re-assembled into sequential order
// See comment on uploadWaitReasons for rationale.
var downloadWaitReasons = []WaitReason{
	// Done by the transfer initiation function (i.e. chunkfunc creation loop)
	EWaitReason.CreateLocalFile(),
	EWaitReason.RAMToSchedule(),

	// Waiting for a work Goroutine to pick up the chunkfunc and execute it.
	// Chunks in this state are effectively a queue of work, waiting for their network downloads to be initiated
	EWaitReason.WorkerGR(),

	// Waiting until the per-file pacer (if any applies to this download) says we can proceed
	EWaitReason.FilePacer(),

	// These next ones are the actual network activity
	EWaitReason.HeaderResponse(),
	EWaitReason.Body(),
	// next two exist, but are not reported on separately in GetCounts, so are commented out
	//EWaitReason.BodyReReadDueToMem(),
	//EWaitReason.BodyReReadDueToSpeed(),

	// Sorting and QueueToWrite together comprise a queue of work waiting to be written to disk.
	// The former are unsorted, and the latter have been sorted into sequential order.
	// PriorChunk is unusual, because chunks in that wait state are not (yet) waiting for their turn to be written to disk,
	// instead they are waiting on some prior chunk to finish arriving over the network
	EWaitReason.Sorting(),
	EWaitReason.PriorChunk(),
	EWaitReason.QueueToWrite(),

	// The actual disk write
	EWaitReason.DiskIO(),

	EWaitReason.Epilogue(),
	// Plus Done/cancelled, which are not included here because not wanted for GetCounts
}

var s2sCopyWaitReasons = []WaitReason{
	// Waiting for a worker Go routine to pick up the scheduled chunk func.
	// Chunks in this state are effectively a queue of work waiting to be sent over the network
	EWaitReason.WorkerGR(),

	// Waiting until the per-file pacer (if any applies to this s2sCopy) says we can proceed
	EWaitReason.FilePacer(),

	// Start to send Put*FromURL, then S2S copy will start in service side, and Azcopy will wait the response which indicates copy get finished.
	EWaitReason.S2SCopyOnWire(),

	EWaitReason.Epilogue(),
}

func (wr WaitReason) String() string {
	return string(wr.Name) // avoiding reflection here, for speed, since will be called a lot
}

type ChunkStatusLogger interface {
	LogChunkStatus(id ChunkID, reason WaitReason)
	IsWaitingOnFinalBodyReads() bool
}

type ChunkStatusLoggerCloser interface {
	ChunkStatusLogger
	GetCounts(td TransferDirection) []chunkStatusCount
	GetPrimaryPerfConstraint(td TransferDirection, rc RetryCounter) PerfConstraint
	FlushLog() // not close, because we had issues with writes coming in after this // TODO: see if that issue still exists
	CloseLogger()
}

type RetryCounter interface {
	GetTotalRetries() int64
}

// chunkStatusLogger records all chunk state transitions, and makes aggregate data immediately available
// for performance diagnostics. Also optionally logs every individual transition to a file.
type chunkStatusLogger struct {
	atomicLastRetryCount            int64
	atomicIsWaitingOnFinalBodyReads int32
	counts                          []int64
	outputEnabled                   bool
	unsavedEntries                  chan *chunkWaitState
	flushDone                       chan struct{}
	cpuMonitor                      CPUMonitor
}

func NewChunkStatusLogger(jobID JobID, cpuMon CPUMonitor, logFileFolder string, enableOutput bool) ChunkStatusLoggerCloser {
	logger := &chunkStatusLogger{
		counts:         make([]int64, numWaitReasons()),
		outputEnabled:  enableOutput,
		unsavedEntries: make(chan *chunkWaitState, 1000000),
		flushDone:      make(chan struct{}),
		cpuMonitor:     cpuMon,
	}
	if enableOutput {
		chunkLogPath := path.Join(logFileFolder, jobID.String()+"-chunks.log") // its a CSV, but using log extension for consistency with other files in the directory
		go logger.main(chunkLogPath)
	}
	return logger
}

func numWaitReasons() int32 {
	return EWaitReason.Cancelled().index + 1 // assume that maintainers follow the comment above to always keep Cancelled as numerically the greatest one
}

type chunkStatusCount struct {
	WaitReason WaitReason
	Count      int64
}

type chunkWaitState struct {
	ChunkID
	reason    WaitReason
	waitStart time.Time
}

////////////////////////////////////  basic functionality //////////////////////////////////

func (csl *chunkStatusLogger) LogChunkStatus(id ChunkID, reason WaitReason) {
	// always update the in-memory stats, even if output is disabled
	csl.countStateTransition(id, reason)

	if !csl.outputEnabled || id.IsPseudoChunk() { // pseudo chunks are only for aggregate stats, not detailed logging
		return
	}

	csl.unsavedEntries <- &chunkWaitState{ChunkID: id, reason: reason, waitStart: time.Now()}
}

func (csl *chunkStatusLogger) FlushLog() {
	if !csl.outputEnabled {
		return
	}

	// In order to be idempotent, we don't close any channel here, we just flush it

	csl.unsavedEntries <- nil // tell writer that it must flush, then wait until it has done so

	<-csl.flushDone
}

// CloseLogger close the chunklogger thread.
func (csl *chunkStatusLogger) CloseLogger() {
	// Once logger is closed, we log no more chunks.
	csl.outputEnabled = false

	/*
	 * No more chunks will ever be written, let the main logger know about this.
	 * On closing this channel the main logger will exit from its for-range loop.
	 */
	close(csl.unsavedEntries)
}

func (csl *chunkStatusLogger) main(chunkLogPath string) {
	f, err := os.Create(chunkLogPath)
	if err != nil {
		panic(err.Error())
	}
	defer func() { _ = f.Close() }()

	w := bufio.NewWriter(f)
	_, _ = w.WriteString("Name,Offset,State,StateStartTime\n")

	doFlush := func() {
		_ = w.Flush()
		_ = f.Sync()
	}
	defer doFlush()

	alwaysFlushFromNowOn := false

	// We will exit the following for-range loop after CloseLogger() closes the csl.unsavedEntries channel.
	for x := range csl.unsavedEntries {
		if x == nil {
			alwaysFlushFromNowOn = true
			doFlush()
			csl.flushDone <- struct{}{}
			continue // TODO can become break (or be moved to later if we close unsaved entries, once we figure out how we got stuff written to us after CloseLog was called)

		}
		_, _ = w.WriteString(fmt.Sprintf("%s,%d,%s,%s\n", x.Name, x.OffsetInFile(), x.reason, x.waitStart))
		if alwaysFlushFromNowOn {
			// TODO: remove when we figure out how we got stuff written to us after CloseLog was called. For now, this should handle those cases (if they still exist)
			doFlush()
		}

	}
}

////////////////////////////// aggregate count and analysis support //////////////////////

// We maintain running totals of how many chunks are in each state.
// To do so, we must determine the new state (which is simply a parameter) and the old state.
// We obtain and track the old state within the chunkID itself. The alternative, of having a threadsafe
// map in the chunkStatusLogger, to track and look up the states, is considered a risk for performance.
func (csl *chunkStatusLogger) countStateTransition(id ChunkID, newReason WaitReason) {

	// NOTE to maintainers: this routine must be idempotent. E.g. for some whole-of-file pseudo chunks,
	// they may be set to "Done" more than once.  So this routine should work OK if status is set to
	// a status that the chunk already has

	// Flip the chunk's state to indicate the new thing that it's waiting for now
	oldReasonIndex := atomic.SwapInt32(id.waitReasonIndex, newReason.index)

	// Update the counts
	// There's no need to lock the array itself. Instead just do atomic operations on the contents.
	// (See https://groups.google.com/forum/#!topic/Golang-nuts/Ud4Dqin2Shc)
	if oldReasonIndex > 0 && oldReasonIndex < int32(len(csl.counts)) {
		atomic.AddInt64(&csl.counts[oldReasonIndex], -1)
	}
	if newReason.index < int32(len(csl.counts)) {
		atomic.AddInt64(&csl.counts[newReason.index], 1)
	}
}

func (csl *chunkStatusLogger) getCount(reason WaitReason) int64 {
	return atomic.LoadInt64(&csl.counts[reason.index])
}

// Gets the current counts of chunks in each wait state
// Intended for performance diagnostics and reporting
func (csl *chunkStatusLogger) GetCounts(td TransferDirection) []chunkStatusCount {
	var allReasons []WaitReason

	switch td {
	case ETransferDirection.Upload():
		allReasons = uploadWaitReasons
	case ETransferDirection.Download():
		allReasons = downloadWaitReasons
	case ETransferDirection.S2SCopy():
		allReasons = s2sCopyWaitReasons
	}

	result := make([]chunkStatusCount, len(allReasons))
	for i, reason := range allReasons {
		count := csl.getCount(reason)

		// for simplicity in consuming the results, all the body read states are rolled into one here
		if reason == EWaitReason.BodyReReadDueToSpeed() || reason == EWaitReason.BodyReReadDueToMem() {
			panic("body re-reads should not be requested in counts. They get rolled into the main Body one")
		}
		if reason == EWaitReason.Body() {
			count += csl.getCount(EWaitReason.BodyReReadDueToSpeed())
			count += csl.getCount(EWaitReason.BodyReReadDueToMem())
		}

		result[i] = chunkStatusCount{reason, count}
	}
	return result
}

func (csl *chunkStatusLogger) GetPrimaryPerfConstraint(td TransferDirection, rc RetryCounter) PerfConstraint {
	newCount := rc.GetTotalRetries()
	oldCount := atomic.SwapInt64(&csl.atomicLastRetryCount, newCount)
	retriesSinceLastCall := newCount - oldCount

	switch {
	// it seems sensible to report file pacer (Service) constraint as a higher priority than Disk, if both exist at the same time (but usually they won't)
	case csl.isConstrainedByFilePacer():
		return EPerfConstraint.PageBlobService() // distinguish this from ordinary service throttling for ease of diagnostic understanding (page blobs have per-blob limits)

	// check this ahead of disk, because for uploads retries can force disk activity, and so can be mistaken as a disk constraint
	// if we looked at disk first
	case retriesSinceLastCall > 0:
		return EPerfConstraint.Service()

	case td == ETransferDirection.Upload() && csl.isUploadDiskConstrained():
		return EPerfConstraint.Disk()

	case td == ETransferDirection.Download() && csl.isDownloadDiskConstrained():
		return EPerfConstraint.Disk()

	case csl.cpuMonitor.CPUContentionExists():
		return EPerfConstraint.CPU()

	default:
		return EPerfConstraint.Unknown()
	}
}

const (
	nearZeroQueueSize = 10 // TODO: is there any intelligent way to set this threshold? It's just an arbitrary guestimate of "small" at the moment
)

func (csl *chunkStatusLogger) isConstrainedByFilePacer() bool {
	haveBigQueueForPacer := csl.getCount(EWaitReason.FilePacer()) >= nearZeroQueueSize
	return haveBigQueueForPacer
}

// is disk the bottleneck in an upload?
func (csl *chunkStatusLogger) isUploadDiskConstrained() bool {
	// If we are uploading, and there's almost nothing waiting to go out over the network, then
	// probably the reason there's not much queued is that the disk is slow.
	// BTW, we can't usefully look at any of the _earlier_ states, because they happen in the _generation_ of the chunk funcs
	// (not the _execution_ and so their counts will just tend to equal that of the small goroutine pool that runs them).
	// It might be convenient if we could compare TWO queue sizes here, as we do in isDownloadDiskConstrained, but unfortunately our
	// Jan 2019 architecture only gives us ONE useful queue-like state when uploading, so we can't compare two.
	queueForNetworkIsSmall := csl.getCount(EWaitReason.WorkerGR()) < nearZeroQueueSize

	beforeGRWaitQueue := csl.getCount(EWaitReason.RAMToSchedule()) + csl.getCount(EWaitReason.DiskIO())
	areStillReadingDisk := beforeGRWaitQueue > 0 // size of queue for network is irrelevant if we are no longer actually reading disk files, and therefore no longer putting anything into the queue for network

	return areStillReadingDisk && queueForNetworkIsSmall
}

// is disk the bottleneck in a download?
func (csl *chunkStatusLogger) isDownloadDiskConstrained() bool {
	// See how many chunks are waiting on the disk. I.e. are queued before the actual disk state.
	// Don't include the "PriorChunk" state, because that's not actually waiting on disk at all, it
	// can mean waiting on network and/or waiting-on-Storage-Service. We don't know which. So we just exclude it from consideration.
	chunksWaitingOnDisk := csl.getCount(EWaitReason.Sorting()) + csl.getCount(EWaitReason.QueueToWrite())

	// i.e. are queued before the actual network states
	chunksQueuedBeforeNetwork := csl.getCount(EWaitReason.WorkerGR())

	// if we have way more stuff waiting on disk than on network, we can assume disk is the bottleneck
	const activeDiskQThreshold = 10
	const bigDifference = 5                                            // TODO: review/tune the arbitrary constant here
	isDiskConstrained := chunksWaitingOnDisk > activeDiskQThreshold && // this test is in case both are near zero, as they would be near the end of the job
		chunksWaitingOnDisk > bigDifference*chunksQueuedBeforeNetwork

	// while we are here... set an indicator of whether we are waiting on body reads (only) with nothing more to download
	// TODO: find a better place for this code
	const finalBodyReadsThreshold = 50 // an empirically-derived guestimate of a suitable value.  Too high, and we trigger the final waiting logic too soon; too low and we trigger to too late
	chunksBeforeBody := csl.getCount(EWaitReason.RAMToSchedule()) + chunksQueuedBeforeNetwork + csl.getCount(EWaitReason.HeaderResponse())
	chunksWaitingOnBody := csl.getCount(EWaitReason.Body())
	isSmallNumberWaitingOnBody := chunksWaitingOnBody > 0 && chunksWaitingOnBody < finalBodyReadsThreshold
	if chunksBeforeBody == 0 && isSmallNumberWaitingOnBody {
		atomic.StoreInt32(&csl.atomicIsWaitingOnFinalBodyReads, 1) // there's nothing BEFORE the body stage, so the body stage is the hold-up
	} else {
		atomic.StoreInt32(&csl.atomicIsWaitingOnFinalBodyReads, 0)
	}

	return isDiskConstrained
}

func (csl *chunkStatusLogger) IsWaitingOnFinalBodyReads() bool {
	return atomic.LoadInt32(&csl.atomicIsWaitingOnFinalBodyReads) == 1 // not computed on demand, because there will be LOTS of calls (>= 1 per chunk)
}

///////////////////////////////////// Sample LinqPad query for manual analysis of chunklog /////////////////////////////////////

/* LinqPad query used to analyze/visualize the CSV as is follows:
   Needs CSV driver for LinqPad to open the CSV - e.g. https://github.com/dobrou/CsvLINQPadDriver

var data = chunkwaitlog_noForcedRetries;

const int assumedMBPerChunk = 8;

DateTime? ParseStart(string s)
{
	const string format = "yyyy-MM-dd HH:mm:ss.fff";
	var s2 = s.Substring(0, format.Length);
	try
	{
		return DateTime.ParseExact(s2, format, CultureInfo.CurrentCulture);
	}
	catch
	{
		return null;
	}
}

// convert to real datetime (default unparsable ones to a fixed value, simply to avoid needing to deal with nulls below, and because all valid records should be parseable. Only exception would be something partially written a time of a crash)
var parsed = data.Select(d => new { d.Name, d.Offset, d.State, StateStartTime = ParseStart(d.StateStartTime) ?? DateTime.MaxValue}).ToList();

var grouped = parsed.GroupBy(c => new {c.Name, c.Offset});

var statesForOffset = grouped.Select(g => new
{
	g.Key,
	States = g.Select(x => new { x.State, x.StateStartTime }).OrderBy(x => x.StateStartTime).ToList()
}).ToList();

var withStatesOfInterest = (from sfo in statesForOffset
let states = sfo.States
let lastIndex = states.Count - 1
let statesWithDurations = states.Select((s, i) => new{ s.State, s.StateStartTime, Duration = ( i == lastIndex ? new TimeSpan(0) : states[i+1].StateStartTime - s.StateStartTime) })
let hasLongBodyRead = statesWithDurations.Any(x => (x.State == "Body" && x.Duration.TotalSeconds > 30)  // detect slowness in tests where we turn off the forced restarts
|| x.State.StartsWith("BodyReRead"))                       // detect slowness where we solved it by a forced restart
select new {sfo.Key, States = statesWithDurations, HasLongBodyRead = hasLongBodyRead})
.ToList();

var filesWithLongBodyReads = withStatesOfInterest.Where(x => x.HasLongBodyRead).Select(x => x.Key.Name).Distinct().ToList();

filesWithLongBodyReads.Count().Dump("Number of files with at least one long chunk read");

var final = (from wsi in withStatesOfInterest
join f in filesWithLongBodyReads on wsi.Key.Name equals f
select new
{
ChunkID = wsi.Key,
wsi.HasLongBodyRead,
wsi.States

})
.GroupBy(f => f.ChunkID.Name)
.Select(g => new {
Name = g.Key,
Chunks = g.Select(x => new {
OffsetNumber = (int)(long.Parse(x.ChunkID.Offset)/(assumedMBPerChunk*1024*1024)),
OffsetValue = x.HasLongBodyRead ? Util.Highlight(x.ChunkID.Offset) : x.ChunkID.Offset, States = x.States}
).OrderBy(x => x.OffsetNumber)
})
.OrderBy(x => x.Name);

final.Dump();

*/